Global data-free area visibility forecasting method considering geographical distance

Abstract

The invention discloses a global no-data area visibility forecasting method considering geographical distance, and relates to the technical field of visibility forecasting, and the global no-data area visibility forecasting method considering geographical distance mainly comprises the steps: carrying out the preprocessing of meteorological grid data and visibility monitoring data, and obtaining a forecasting data set; the constructed artificial intelligence model comprises an encoder and a decoder; and training and evaluating the artificial intelligence model by using the forecast data set to obtain a basic artificial intelligence model, and predicting the visibility of the global data-free area to obtain a prediction result. By implementing the global data-free area visibility forecasting method considering the geographic distance, the seasonal forecasting accuracy of the visibility of the global station-free area can be improved.

Description

Technical Field

[0001] This invention relates to the field of visibility forecasting technology, and more specifically, to a method for visibility forecasting in globally unrecorded areas that takes into account geographical distance. Background Technology

[0002] Visibility is a crucial indicator of atmospheric transparency. As a key physical quantity related to air pollution, visibility is extremely sensitive to changes in environmental and meteorological conditions, and its fluctuations are a significant cause of accidents in aerospace, maritime, and land transportation. Changes in visibility have already posed a serious threat to human life and property. Especially against the backdrop of escalating air pollution, the potential risks posed by changes in visibility to socio-economic activities and human health are even more pronounced. Efficient and accurate early warning systems for changes in visibility are essential for mitigating these risks.

[0003] Countries such as the UK, the US, and China have successively established ground monitoring stations to obtain high-precision ground visibility data. However, constructing these stations requires significant investment of manpower and resources, and their maintenance and operation also require long-term, continuous investment. Furthermore, the monitoring stations are sparsely distributed, mainly concentrated in urban areas, and are constrained by geographical and temporal conditions, especially in special geographical environments such as deserts, glaciers, and polar regions, further increasing the difficulty of station establishment. On the other hand, non-monitoring areas cannot effectively monitor visibility changes, making it difficult to identify high-risk periods such as fog seasons and sandstorms in advance, posing significant challenges to the safety and efficiency of agricultural production, urban planning, and other fields. Therefore, achieving accurate visibility forecasting in non-monitoring areas is an important task.

[0004] The spatial correlation derived from the first law of geography can be used to characterize the geographical features of a region, including location information, topographic information, and similar visibility changes. Considering the role of geographical distance information has a very important impact on the accurate forecasting of visibility.

[0005] Artificial intelligence (AI) technology can automatically extract complex nonlinear features from massive amounts of meteorological data, reducing reliance on subjective experience. Compared with traditional numerical weather prediction and statistical forecasting methods, AI technology shows greater potential in predicting complex weather sequences. AI can accurately fit and forecast visibility information from various auxiliary data. Therefore, effectively combining geographic distance information with AI technology not only provides a solid technical foundation for proactive disaster forecasting during high-risk periods, but also enables AI to learn more geographic information, effectively improving the accuracy of visibility forecasts.

[0006] While many current artificial intelligence techniques have achieved relatively accurate near-term and short-term visibility forecasts, significant uncertainties remain regarding seasonal visibility warnings, and the accuracy of seasonal forecasts needs further evaluation. Currently, the following issues exist with global visibility forecasting: (1) Short-term visibility forecasts mainly focus on transient changes and often fail to capture high-risk periods related to seasonal disaster events, thus having inherent limitations.

[0007] (2) Seasonal changes in visibility are a complex process influenced by multiple factors. Against the backdrop of global warming, the frequency and intensity of extreme weather events are constantly increasing, disrupting previously stable climate patterns and altering traditional visibility trends. This increased complexity brings significant uncertainty to accurately simulating the dynamic relationship between climate change and visibility.

[0008] (3) The spatial distribution of global weather stations is significantly uneven. Current progress in visibility forecasting relies heavily on data from existing ground monitoring stations, resulting in considerable uncertainty in areas without such stations.

[0009] (4) The visibility of adjacent meteorological monitoring stations is correlated. If only meteorological elements are considered, the artificial intelligence will lack the ability to learn geographical information. The role of geographical distance information in improving the forecast effect needs to be evaluated.

[0010] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0011] The purpose of this invention is to provide a visibility forecasting method for globally data-free areas that takes into account geographical distance, thereby improving the accuracy of seasonal visibility forecasts for globally data-free areas.

[0012] This invention provides a method for visibility forecasting in globally unrecorded areas that takes into account geographical distance, comprising the following steps: S1: Preprocess meteorological raster data and visibility monitoring data to obtain a forecast dataset; S2: Construct an artificial intelligence model, which includes an encoder and a decoder; S3: Train the artificial intelligence model using the forecast dataset, and evaluate the overall forecast performance of the trained artificial intelligence model using the coefficient of determination, mean absolute error, and root mean square error to obtain a basic artificial intelligence model; use the basic artificial intelligence model to predict visibility in areas with no global data to obtain the prediction results.

[0013] Furthermore, the aforementioned methods for visibility forecasting in globally unrecorded areas that take geographical distance into account also include: S4: Based on the visibility monitoring data, calculate the geographical distance to obtain the geographical distance between monitoring stations; S5: The artificial intelligence model is trained using the geographical distance between the forecast dataset and the monitoring station, and the overall forecast performance of the trained artificial intelligence model is evaluated using the coefficient of determination, mean absolute error, and root mean square error to obtain an artificial intelligence model that takes geographical distance into account; the visibility of data-free areas around the world is predicted using the artificial intelligence model that takes geographical distance into account to obtain the prediction results.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for predicting visibility in globally data-free areas taking into account geographical distance.

[0015] Implementing the global visibility forecasting method for data-free areas that takes geographical distance into account, as provided in this invention, has the following beneficial effects: This invention addresses the problems of errors, uncertainties, insufficient consideration of auxiliary information, and difficulty in accurately forecasting seasonal changes in visibility in areas without monitoring stations in previous visibility forecasts. It acquires meteorological reanalysis data of different dimensions and visibility data from globally available ground monitoring stations; it matches reanalysis data with the same temporal and spatial resolution based on the geographical location of the ground-based monitoring stations; it constructs an autocorrelation of visibility with geographical distance based on the first law of geography and the Harvard-Thailand formula; it forecasts visibility changes over the next 12 months using a bidirectional encoder-decoder structure; and it introduces geographical distance into visibility forecasting, enabling more accurate 12-month visibility forecasts for areas without ground-based monitoring stations.

[0016] This invention, through an encoder-decoder architecture, can output multi-step forecast results based on the historical sequence of input data. The encoder, employing a BiLSTM, can fully utilize historical, future, and current auxiliary information to capture the temporal relationship between auxiliary data and visibility information. Re-concatenating the hidden state and cell state and inputting them into the LSTM (Long Short-Term Memory) layer not only improves data utilization but also effectively fits the nonlinear relationship between auxiliary data and visibility.

[0017] This invention improves the fitting ability of auxiliary data and visibility data by adding a nonlinear hidden layer and a learning rate adjustment strategy to the forecast model, enabling the model to quickly and accurately find the optimal solution and improve the accuracy of the forecast.

[0018] This invention utilizes the autocorrelation between visibility data monitored at the current station and that of neighboring stations. Based on the first law of geography, this geographical distance information is numerically represented and incorporated into the forecasting model. This enables the model to learn more geographical location information and spatial autocorrelation, thereby improving the accuracy and universality of the model's forecasts.

[0019] This invention overcomes the inherent limitations of short-term and near-term forecasts, effectively considers the role of geographical distance information, integrates artificial intelligence technology with the first law of geography, and more accurately realizes seasonal forecasts of visibility in global no-site areas. Attached Figure Description The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the global visibility forecasting method for data-free areas that takes geographical distance into account, provided by the present invention. Figure 2 This is a schematic diagram of the technical solution of the global data-free area visibility forecasting method that takes geographical distance into account, provided by the present invention. Figure 3 This is a diagram of the artificial intelligence model architecture of the present invention.

[0020] Figure 4 This is a schematic diagram of the data used for model training provided by the present invention; Figure 5 This is a schematic diagram of the original ground-measured data provided by the present invention; Figure 6 This is a schematic diagram of the forecast error range and proportion provided by the present invention. Detailed Implementation

[0021] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] Figure 1 A schematic diagram of the visibility forecasting method for globally unrecorded areas taking geographical distance into account in this embodiment is shown. In this embodiment, the visibility forecasting method for globally unrecorded areas taking geographical distance into account includes the following steps: S1: Preprocess meteorological raster data and visibility monitoring data to obtain a forecast dataset; As an exemplary embodiment, in step S1, meteorological grid data with different dimensions and the actual visibility data of ground monitoring stations are acquired, and the raw data are processed to unify the units. Specifically, the data mentioned above in step S1 includes: ERA5-Land reanalysis data and visibility monitoring data provided by NCEI under NOAA; In one exemplary embodiment, the preprocessing includes: unifying the temporal and spatial resolution of meteorological raster data based on the temporal resolution of visibility at ground-based monitoring stations; extracting the pixel values ​​of the corresponding raster data based on the specific latitude and longitude of the monitoring stations, removing outliers, and obtaining a forecast dataset. S2: Construct an artificial intelligence model, which includes an encoder and a decoder; In one exemplary embodiment, the encoder is configured to: compress meteorological and geographical distance data into a fixed vector based on a bidirectional long short-term memory neural network, and capture dependencies and trends within auxiliary data and visibility at each time step; In one exemplary embodiment, the encoder includes a forward propagation LSTM layer and a backward propagation LSTM layer; The forward propagation LSTM layer is used to receive the current input meteorological and geographical distance information, and to filter and update the information through a gating mechanism, thereby capturing the temporal characteristics of meteorological variables and fusing geographical distance and meteorological conditions. The backpropagation LSTM layer is used to track the impact of auxiliary data on visibility changes starting from the end of the input sequence and to fill in necessary information that the forward propagation LSTM layer may have missed. In one exemplary embodiment, the encoder and the decoder are connected via a Concatenate function, which is used to reassemble the states of the hidden layer and the cells. As an exemplary embodiment, the encoder employs a bidirectional long short-term network (BiLSTM) structure with 256 internal input neurons, incorporating forward and backward hidden layers and cell states to adapt to the encoder's input. In one exemplary embodiment, the decoder uses the RepeatVector function to copy the final state vector of the encoder to obtain a three-dimensional tensor that conforms to the decoder input. As an exemplary embodiment, the decoder internally inputs... One neuron; In one exemplary embodiment, the artificial intelligence model further includes a batch normalization layer, a non-linear activation function, and a gradient descent learning rate adjuster; S3: Train the artificial intelligence model using the forecast dataset, and evaluate the overall forecast performance of the trained artificial intelligence model using the coefficient of determination, mean absolute error, and root mean square error to obtain a basic artificial intelligence model; use the basic artificial intelligence model to predict visibility in areas with no global data to obtain the prediction results.

[0023] In one exemplary embodiment, the visibility forecasting method for globally unrecorded areas that takes into account geographical distance further includes: S4: Based on the visibility monitoring data, calculate the geographical distance to obtain the geographical distance between monitoring stations; In one exemplary embodiment, step S4 specifically includes: calculating the geographical distance based on the visibility monitoring data to obtain the geographical distance between monitoring stations, as shown in the formula: , , , , , in, Numerical expression for distances on the Earth's spherical surface; and These represent the latitude and longitude differences between given monitoring points; and Indicates the latitude of the two stations; The central angle is the line connecting two points on the Earth's surface to the center of the sphere. For a certain point and point Spatial distance between them; The radius of the Earth; This provides visibility information for all regions and geographical distance to the current observation point. For the first The weights of each sample point; Indicates at point Observed visibility data; S5: The artificial intelligence model is trained using the geographical distance between the forecast dataset and the monitoring station, and the overall forecast performance of the trained artificial intelligence model is evaluated using the coefficient of determination, mean absolute error, and root mean square error to obtain an artificial intelligence model that takes geographical distance into account; the visibility of data-free areas around the world is predicted using the artificial intelligence model that takes geographical distance into account to obtain the prediction results.

[0024] In some embodiments, the above-described method for predicting visibility in globally unrecorded areas, taking into account geographical distance, can also be implemented in the following ways.

[0025] In this embodiment, the visibility forecasting method for globally unrecorded areas taking into account geographical distance includes the following steps: Step 1: Acquire meteorological grid data with different dimensions and actual visibility data from ground-based monitoring stations; The different data in step 1 include: ERA5-Land reanalysis data and visibility monitoring data provided by NCEI under NOAA.

[0026] Step 2: Unify the spatiotemporal resolution of visibility and meteorological data to construct a forecast dataset; The specific process in step 2 is as follows: Step 21: Based on the temporal resolution of visibility at ground-based monitoring stations, standardize the temporal resolution of meteorological raster data. Standardize the spatial resolution of raster data.

[0027] Step 22: Extract the pixel values ​​of the corresponding raster data based on the specific latitude and longitude of the monitoring station.

[0028] Step 23: Remove outliers from the extracted data, clean the data, and generate a forecast dataset.

[0029] Step 3: Construct a formula for calculating geographical distance to obtain the geographical distance between monitoring stations; The specific process in step 3 is as follows: Step 31: Based on the distance characteristics of things according to the first law of geography, take the reciprocal and square of the distance between monitoring stations to highlight the influence of distance factors.

[0030] Step 32: Since the Earth is an ellipsoid, the Haverthyn formula is used to calculate the arc distance between two points.

[0031] Step 33: Using steps 31 and 32, calculate the geographical distance between a point and all stations to obtain the specific value of the geographical distance.

[0032] Step 4: Forecast visibility values ​​for the next 12 months using an encoder-decoder approach; The specific process of constructing the AI ​​model in step 4 is as follows: Step 41: Construct the encoder using a bidirectional long short-term network structure (BiLSTM) with 256 input neurons.

[0033] Step 42: Merge the forward and backward hidden layers and cell states to adapt to the encoder input.

[0034] Step 43: Construct a decoder, which has 256*2 neurons input to improve the value of information utilization.

[0035] Step 44: Add batch normalization, non-linear activation function, and gradient descent learning rate adjuster to improve prediction accuracy.

[0036] Step 5: Incorporate geographical distance to improve forecast accuracy.

[0037] Step 5 mainly involves adding the geographical distance factor constructed in step 3 to the forecast in step 4, forming a visibility forecast that takes geographical distance into account, and obtaining more reliable forecast results.

[0038] In some embodiments, the above-described method for predicting visibility in globally unrecorded areas, taking into account geographical distance, can also be implemented in the following ways.

[0039] Please refer to Figure 2 , Figure 2 This is a flowchart of the method in this embodiment. In this embodiment, the visibility forecasting method for globally unrecorded areas taking into account geographical distance includes the following steps: Step 1: Acquire meteorological grid data with different dimensions and actual visibility data from ground-based monitoring stations; The meteorological raster data used in this embodiment is sourced from the European Centre for Medium-Range Weather Forecasts' Generation 5 Land Surface Reanalysis Dataset (ERA5-Land). The meteorological data names and dimensions are referenced in Table 1. The measured visibility data is sourced from the National Center for Environmental Information (NCEI) under the National Oceanic and Atmospheric Administration (NOAA). Continuous monitoring stations from 1995 to 2023 were selected as training data stations. The original data provided visibility units in miles, which were converted to kilometers to conform to international conventions.

[0040] Table 1: Meteorological Raster Data

[0041] Step 2: Unify the spatiotemporal resolution of visibility and meteorological data to construct a forecast dataset; It should be noted that the specific process of constructing the dataset in step S2 is as follows: (i) The raw visibility monitoring data provides a time resolution of hourly and daily. Due to omissions and anomalies caused by instruments and natural phenomena, stations with 5 consecutive days of missing data and 30 days of missing data throughout the year are removed. Ground monitoring points with continuous time series and high-quality data are selected, and the monitoring data are averaged to generate monthly ground monitoring data.

[0042] (ii) The ERA5-Land has a monthly time resolution, which is the same as the quality-controlled ground monitoring data, and can be directly used for processing.

[0043] (iii) Based on the latitude and longitude of the ground monitoring stations after quality control, extract the pixel values ​​of the corresponding meteorological variables in the ERA5-Land data to form a forecast dataset.

[0044] Step 3: Construct a geographical distance calculation formula to obtain geographical distance information between monitoring stations; The specific process of constructing geographic distance information in step S3 is as follows: (i) For visibility, atmospheric and meteorological conditions in neighboring locations are often similar, leading to strong correlations between their visibility levels. Incorporating geographic distance into artificial intelligence models allows them to utilize these spatial correlations. When predicting visibility at a selected location, the model can reference visibility conditions and related characteristics at nearby and distant locations. Globally, visibility typically exhibits gradient variations due to geographic differences. Using geographic distance as an independent variable enables artificial intelligence to capture these spatial gradients, thereby improving its ability to predict visibility trends at different locations. By integrating spatial relationships and visibility data at different distances, artificial intelligence models can uncover hidden patterns, thereby improving the accuracy and reliability of forecasts.

[0045] (ii) Since the Earth is approximately spherical, surface arc distances are used instead of simple linear distances to calculate the distances between locations. This embodiment uses the Haversine formula to calculate the distances between monitoring stations.

[0046]

[0047]

[0048]

[0049] in and These represent the latitude and longitude differences between the given monitoring points. and This indicates the latitude of the two stations. It is the Earth's radius, approximately 6371 kilometers. For a certain point and point The spatial distance between them.

[0050] (iii) Based on the first law of geography, similar things are highly correlated and the greater the distance, the weaker the correlation. In this embodiment, the reciprocal and square of the distance are taken in the formula to approximate the properties of the geography law.

[0051]

[0052]

[0053] Indicates at point Observed visibility data. GD represents the visibility across all areas and the geographic distance to the current observation point.

[0054] Step 4: Forecast visibility values ​​for the next 12 months using an encoder-decoder approach; The specific process of AI model prediction in step S4 is as follows: (I) This embodiment develops an encoder-decoder model, where the encoder, based on a bidirectional long short-term memory neural network (BiLSTM), compresses meteorological and geographical distance data into a fixed vector and captures the dependencies and trends within visibility at each time step. The decoder receives two merged states (hidden state and cell state) generated by the encoder, enabling the decoder to make forecasts based on historical information extracted from the encoder, thereby generating a visibility sequence from January to December, such as... Figure 3 The diagram shown is a structural schematic of an artificial intelligence model. (ii) Input data: The meteorological historical sequence (results extracted from ERA5-Land) and the geographical distance historical sequence are run on a BiLSTM, including different times (0-n), and the output generates visibility forecast results for 1-12 months.

[0055] (iii) Forward refers to the forward propagation LSTM layer, which receives the current input meteorological and geographical distance information and filters and updates the information through gating mechanisms (forget gate, input gate, and output gate) to capture the temporal characteristics of meteorological variables and fuse geographical distance and meteorological conditions. Backward is the backward propagation LSTM layer, which tracks the impact of auxiliary data on visibility changes starting from the end of the input sequence and fills in necessary information that may have been missed by the forward layer.

[0056] (iv) Concatenate is used to reassemble the hidden layer and cell states. BiLSTM captures bidirectional dependencies in time-series data by simultaneously processing the forward and reverse order of the input sequence. For each time step, it generates two hidden states (1. forward hidden state, 2. backward hidden state). The generation of these two states requires the decoder to utilize information from both directions simultaneously to generate accurate visibility forecasts, doubling the dimensionality of the merged state. In this embodiment, RepeatVector is used to copy the encoder's final state vector, forming a three-dimensional tensor that conforms to the decoder's input. This implicitly allows the decoder to reference the complete input information when generating the output at each time step. This concatenation method can effectively capture the long-term trends and periodicity of auxiliary data and is compatible with the decoder's decoding process, thereby improving the overall forecast capability.

[0057] (V) Decoder connects batch normalization layer and nonlinear hidden state transition network. Batch normalization layer reduces internal covariate bias by standardizing the mean and variance of input data, making network training more stable and convergent faster. It also reduces overfitting to some extent and improves the model's generalization ability. Adding batch normalization after the output of the decoder LSTM ensures that the features received by subsequent layers have a stable distribution, which helps maintain the stability of predictions. Nonlinear hidden state transition network is used to ensure that the network does not degenerate into a single-layer linear model, and maps the decoder output to a nonlinear space through the tanh activation function, helping the model learn complex data distribution patterns and better capturing continuous changes in time series data. Step 5: Evaluation of the forecasting performance of the artificial intelligence model developed in this embodiment (i) Evaluation metrics are needed to quantify the accuracy of AI model forecasts, and different evaluation metrics have different meanings for the forecast results. The coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE) are used to evaluate the overall forecast performance of AI models.

[0058]

[0059]

[0060]

[0061] Where n represents the total number of visibility samples used for verification. and Let represent the observed and predicted visibility values ​​at time i, respectively. This represents the average of the visibility observations.

[0062] (II) The forecasting model developed in this embodiment needs to be evaluated at different regional scales. Based on longitude and latitude, the forecasting model is applied to the seven continents and different climate zones. The data used for model training is the same as the original ground-based measured data. Figure 4 and Figure 5 As shown.

[0063] (iii) R 2 When MAE and RMSE are applied to seasonal visibility forecasts at different spatial scales, they cannot capture and visualize subtle differences between AI models. To address this limitation, this embodiment employs Relative Error and Relative Bias (RERB) as a robust metric for evaluating the performance of AI models at different regional scales.

[0064]

[0065]

[0066]

[0067] BAI stands for Artificial Intelligence Model that does not consider geographical distance information, while GDAI stands for Visual Distance Model that incorporates geographical distance information into Artificial Intelligence for seasonal visibility forecasting.

[0068] Seasonal forecasts were performed globally for non-site areas using the artificial intelligence model and a basic model that did not consider geographical distance information, based on evaluation metrics. The effectiveness of this invention in visibility forecasting was also verified across intercontinental continents and climate zones. The accuracy of the invention's forecasts for non-site locations was systematically evaluated (based on the difference between forecast and actual values), and the spatial pattern of forecast errors was analyzed, thereby achieving more accurate visibility forecasts.

[0069] Invention verification: In step one, the forecast error range and proportion of this invention at different observation points and different lead times are calculated. More than 1200 stations have forecast errors of less than 1 kilometer, accounting for more than 32% of the total. In the 3rd, 4th, 5th, 8th, 9th, and 10th months, the number of stations with forecast errors less than 1 kilometer exceeds 1450, accounting for more than 38% of the total number of stations. For example... Figure 6 The diagram shown illustrates the forecast error range and proportion. In step two, GDAI showed good average lead time across the three indicators, with an R² of 0.7813, an RMSE of 3.6301, and a MAE of 13.1837. Compared to the baseline model outside this invention, R² increased by 23.06%, while RMSE and MAE decreased by 22.63% and 40.11%, respectively. This invention achieved the highest prediction accuracy within a one-month lead time, with an R² of 0.8017, an RMSE of 3.4579, and a MAE of 11.9573, as detailed in Table 2. Table 2: Performance Comparison of the Two Models

[0070] This embodiment provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for predicting visibility in globally data-free areas while taking into account geographical distance.

[0071] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for predicting visibility in globally unrecorded areas, taking into account geographical distance, characterized in that, Includes the following steps: S1: Preprocess meteorological raster data and visibility monitoring data to obtain a forecast dataset; S2: Construct an artificial intelligence model, which includes an encoder and a decoder; S3: Train the artificial intelligence model using the forecast dataset, and evaluate the overall forecast performance of the trained artificial intelligence model using the coefficient of determination, mean absolute error, and root mean square error to obtain a basic artificial intelligence model; use the basic artificial intelligence model to predict visibility in areas with no global data to obtain the prediction results.

2. The visibility forecasting method for globally unrecorded areas taking into account geographical distance as described in claim 1, characterized in that, The visibility forecasting method for globally undataable areas that takes into account geographical distance also includes: S4: Based on the visibility monitoring data, calculate the geographical distance to obtain the geographical distance between monitoring stations; S5: The artificial intelligence model is trained using the geographical distance between the forecast dataset and the monitoring station, and the overall forecast performance of the trained artificial intelligence model is evaluated using the coefficient of determination, mean absolute error, and root mean square error to obtain an artificial intelligence model that takes geographical distance into account; the visibility of data-free areas around the world is predicted using the artificial intelligence model that takes geographical distance into account to obtain the prediction results.

3. The global visibility forecasting method for data-free areas taking into account geographical distance, as described in claim 1 or 2, is characterized in that... The preprocessing includes: unifying the temporal and spatial resolution of meteorological raster data based on the temporal resolution of visibility at ground-based monitoring stations; extracting the pixel values ​​of the corresponding raster data based on the specific latitude and longitude of the monitoring stations, removing outliers, and obtaining the forecast dataset.

4. The global visibility forecasting method for data-free areas taking into account geographical distance, as described in claim 1 or 2, is characterized in that... The encoder is configured to compress meteorological and geographical distance data into a fixed vector based on a bidirectional long short-term memory neural network, and capture the dependencies and trends within auxiliary data and visibility at each time step.

5. The global visibility forecasting method for data-free areas taking into account geographical distance, as described in claim 1 or 2, is characterized in that... The encoder includes a forward propagation LSTM layer and a backward propagation LSTM layer. The forward propagation LSTM layer is used to receive the current input meteorological and geographical distance information, and to filter and update the information through a gating mechanism, thereby capturing the temporal characteristics of meteorological variables and fusing geographical distance and meteorological conditions. The backward propagation LSTM layer is used to track the impact of auxiliary data on visibility changes starting from the end of the input sequence, and to fill in necessary information that the forward propagation LSTM layer may have missed.

6. The global visibility forecasting method for data-free areas taking into account geographical distance, as described in claim 1 or 2, is characterized in that... The encoder and the decoder are connected via the Concatenate function, which is used to reassemble the hidden layer and the state of the cells.

7. The global visibility forecasting method for data-free areas taking into account geographical distance, as described in claim 1 or 2, is characterized in that... The decoder uses the RepeatVector function to copy the final state vector of the encoder, obtaining a three-dimensional tensor that matches the decoder input.

8. The global visibility forecasting method for data-free areas taking into account geographical distance, as described in claim 1 or 2, is characterized in that... The artificial intelligence model also includes a batch normalization layer, a non-linear activation function, and a gradient descent learning rate adjuster.

9. The global visibility forecasting method for data-free areas taking into account geographical distance as described in claim 2, characterized in that, Step S4 specifically includes: calculating the geographical distance based on the visibility monitoring data to obtain the geographical distance between monitoring stations, as shown in the formula: ， ， ， ， ， in, Numerical expression for distances on the Earth's spherical surface; and These represent the latitude and longitude differences between given monitoring points; and Indicates the latitude of the two stations; The central angle is the line connecting two points on the Earth's surface to the center of the sphere. For a certain point and point Spatial distance between them; The radius of the Earth; This provides visibility information for all regions and geographical distance to the current observation point. For the first The weights of each sample point; Indicates at point Observed visibility data.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the global visibility forecasting method for dataless areas taking into account geographical distance as described in any one of claims 1-9.

Images