ZWD forecasting method and device based on random forest algorithm

Abstract

The invention discloses a ZWD forecasting method and device based on a random forest algorithm, and the method comprises the steps: collecting troposphere delay site data in a global range, carrying out the preprocessing, and extracting time, troposphere dry delay and troposphere wet delay features from the preprocessed troposphere delay site data; taking time and troposphere dry delay characteristics as input and troposphere wet delay characteristics as output, and constructing and training a random forest algorithm model; and finally, performing ZWD forecasting on a to-be-tested area by using the trained random forest algorithm model. The random forest algorithm effectively solves the problem that the global long-time prediction model is too large in processing data volume, prevents the problem of overfitting, and is high in prediction result precision. According to the method, the model is established by adopting observation data of all points in the world within three years, and the troposphere moisture delay of the world in the future one month is forecasted, so that a thought and a method are provided for the long-time troposphere moisture delay model of the world.

Description

Technical Field

[0001] The present invention relates to the field of space geodesy technology, and more specifically, to a ZWD prediction method and device based on a random forest algorithm. Background Art

[0002] Due to the uneven distribution of the troposphere, a certain delay error will occur when the GNSS signal passes through the troposphere. Research usually focuses on the tropospheric zenith delay, including the tropospheric wet delay (ZenithWet Delay, ZWD) caused by water vapor. Although ZWD only accounts for about 10%, the amplitude and range of water vapor changes in the atmosphere are difficult to predict, resulting in low accuracy of the ZWD forecast model.

[0003] Accurate ZWD forecast is of great significance to the development and progress of GNSS technology. The GNSS system determines the position and time of the receiver based on the reception and measurement of satellite signals, and the ZWD in the atmosphere will have a certain impact on the transmission and reception of signals. Accurate estimation of ZWD can correct the signal propagation delay caused by atmospheric humidity, thereby improving the performance and reliability of the communication system. On the other hand, ZWD is also an important data source in GNSS meteorology. As an indicator of atmospheric humidity, ZWD plays an important role in weather forecasting and climate research. Accurate estimation of zenith tropospheric wet delay is conducive to the development of GNSS-related fields such as long-distance communication, navigation, positioning, and weather forecasting. Therefore, it is necessary to study the ZWD prediction model.

[0004] With the efforts of many researchers over the years, many models of tropospheric wet delay have been established. In terms of forecast range, most ZWD forecast models need to predict ZWD based on meteorological data. However, some areas do not have enough meteorological observation stations or do not have timely updated meteorological data, which limits the forecast range of the model. Under complex terrain and geographical conditions, some models cannot fully consider the influence of terrain factors, resulting in low ZWD forecast accuracy in the model for the region. For example, the Bar-Sever model, as a ZWD forecast model based on global meteorological data, can provide useful forecast information in most regions and general meteorological conditions. However, the forecast accuracy is low in mountainous areas and needs further optimization and adaptation, or combined with other terrain influence models to improve the accuracy and reliability of the forecast. The Saastamoinen model is a classic atmospheric delay model, but its scope of application is only in low-latitude areas. Similarly, the forecast range of some empirical models and ZWD forecast models in specific regions is also limited, and there are relatively few global-scale ZWD forecast models. In terms of forecast duration, the change of ZWD is affected by many factors, including weather changes, terrain influences, seasonal changes, etc. This complexity makes it difficult to accurately predict ZWD in the long term. In addition, long-term ZWD prediction requires accurate and continuous observation data, but these data are limited, especially for long-term data records for the world or specific regions. For example, the ZWD prediction model established by Behrend D et al. for the Spanish region, and the prediction model of R. Heinkelmann et al. for estimating ZTD and ZWD using VLBI observation data can only make short-term forecasts. In addition, most empirical models have good forecasting effects in specific time periods, but their prediction capabilities on long time scales are usually poor. In general, most of the existing ZWD forecast models are difficult to make global and long-term forecasts. Summary of the invention

[0005] The present invention proposes a ZWD forecasting method based on a random forest algorithm. Before establishing the model, the data is processed into a time series format, and a global distribution map and a time series map of tropospheric wet delay are plotted, which intuitively shows the changes in tropospheric wet delay in different regions and at different times. When establishing the model, tropospheric dry delay, year, month, day, and annual accumulation of days are used as features, tropospheric wet delay is used as the prediction object, and a random forest algorithm is used for training. After the model is established, a test set prediction result graph, a test set regression graph, a test set error histogram, a feature importance graph, and an error curve graph are plotted, and accuracy indicators such as the sum of squared errors, mean absolute error, mean square error, and root mean square error are calculated to comprehensively evaluate the accuracy of the model, thereby achieving a large-scale and long-term ZWD forecast and improving the accuracy of the forecast.

[0006] To achieve the above object, the present invention provides a ZWD prediction method based on a random forest algorithm in a first aspect, comprising:

[0007] Collect tropospheric delay site data worldwide;

[0008] Preprocess the collected tropospheric delay station data, including annual accumulation and day conversion and data cleaning;

[0009] Extract the time, tropospheric dry delay and tropospheric wet delay characteristics from the preprocessed tropospheric delay station data;

[0010] The random forest algorithm model is constructed and trained by taking time and tropospheric dry delay characteristics as input and tropospheric wet delay characteristics as output;

[0011] The trained random forest algorithm model is used to predict ZWD in the test area.

[0012] In one embodiment, preprocessing the collected tropospheric delay site data includes:

[0013] The collected tropospheric delay station data are converted from year to day, and the original data format is converted into the data format of features and targets;

[0014] Delete incorrect and irrelevant data in the format converted data, and complete incomplete data.

[0015] In one embodiment, the preprocessed tropospheric delay site data includes time series-site data and site-time series data. After preprocessing the collected tropospheric delay site data, the method further includes:

[0016] Draw a spatial distribution map based on time series-site data and analyze the spatial distribution pattern;

[0017] Draw a time series graph based on the site-time series data and analyze the temporal variation trend.

[0018] In one embodiment, extracting time, tropospheric dry delay, and tropospheric wet delay characteristics from preprocessed tropospheric delay station data includes:

[0019] Extract time data from the preprocessed tropospheric delay station data and divide the time into months, days, annual cumulative days and moments;

[0020] Tropospheric dry delay and tropospheric wet delay are extracted from the preprocessed tropospheric delay station data.

[0021] In one embodiment, before taking the time and tropospheric dry delay features as input and the tropospheric wet delay features as output, and constructing and training the random forest algorithm model, the method also includes: performing feature correlation analysis on the extracted time and tropospheric dry delay, and selecting features that meet preset conditions as input.

[0022] In one embodiment, the time and the tropospheric dry delay characteristics are used as inputs, and the tropospheric wet delay characteristics are used as outputs to construct and train a random forest algorithm model, including:

[0023] Construct the original sample set based on the selected features;

[0024] Samples are extracted from the original sample set with replacement to construct multiple different new sample sets. Decision trees are independently constructed based on these new sample sets to form a random forest containing many decision trees. When new samples are added, each decision tree in the random forest independently predicts them and gives corresponding prediction results. Finally, the prediction results of multiple decision trees are integrated to obtain the final prediction result of the model.

[0025] In one embodiment, the method further comprises:

[0026] The accuracy of the constructed random forest algorithm model was tested.

[0027] Based on the same inventive concept, the second aspect of the present invention provides a ZWD forecasting device based on a random forest algorithm, comprising:

[0028] Data collection module, used to collect tropospheric delay site data worldwide;

[0029] Data preprocessing module, used to preprocess the collected tropospheric delay site data, including annual accumulation and day conversion and data cleaning;

[0030] A feature extraction module is used to extract time, tropospheric dry delay and tropospheric wet delay features from the preprocessed tropospheric delay station data;

[0031] A model building module is used to build and train a random forest algorithm model by taking time and tropospheric dry delay characteristics as inputs and tropospheric wet delay characteristics as outputs;

[0032] The ZWD forecast module is used to use the trained random forest algorithm model to perform ZWD forecasts on the test area.

[0033] Based on the same inventive concept, the third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the ZWD prediction method based on the random forest algorithm described in the first aspect.

[0034] Based on the same inventive concept, the fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the ZWD prediction method based on the random forest algorithm described in the first aspect is implemented.

[0035] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:

[0036] The present invention provides a ZWD forecasting method based on a random forest algorithm. First, tropospheric delay site data is collected worldwide; then, pre-processing of annual-to-day conversion and data cleaning is performed, and then time, tropospheric dry delay and tropospheric wet delay features are extracted therefrom; then, time and tropospheric dry delay features are used as inputs, and tropospheric wet delay features are used as outputs to construct and train a random forest algorithm model; finally, the trained random forest algorithm model is used to forecast ZWD for the area to be measured. The present invention uses tropospheric dry delay and time as features, and time is further divided into features such as year, month, day, and annual-to-day. The reason why tropospheric dry delay is used as a feature is that both tropospheric dry delay and tropospheric wet delay are affected by surface temperature and humidity, and tropospheric dry delay is easier to obtain and the measurement data is more accurate. Therefore, the accuracy of ZHD can be used to improve the estimation of ZWD, thereby improving the accuracy of prediction. On the other hand, a random forest algorithm is used to analyze and model three years of global data, and its changing law is explored through machine learning without relying on meteorological data. The random forest algorithm can efficiently process large amounts of data, has fast convergence speed, and strong prediction timeliness. Its excellent prediction accuracy and ability to handle nonlinear relationships have attracted much attention. By performing methods such as random sampling, the algorithm can effectively suppress the overfitting problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0038] Figure 1 is an overall flow chart of a ZWD prediction method based on a random forest algorithm in an embodiment of the present invention;

[0039] Figure 2 is a detailed flow chart of a ZWD prediction method based on a random forest algorithm in an embodiment of the present invention;

[0040] Figure 3 A schematic diagram of sources of data collected in an embodiment of the present invention;

[0041] Figure 4 is a schematic diagram of a random forest algorithm in an embodiment of the present invention;

[0042] Figure 5 It is a heat map of feature correlation matrix in an embodiment of the present invention;

[0043] Figure 6 It is a global distribution map of ZWD in an embodiment of the present invention, wherein (a), (b), (c), and (d) represent the global distribution maps of ZWD on March 1, June 1, September 1, and December 1, 2021, respectively; (e), (f), (g), and (h) represent the global distribution maps of ZWD on March 1, June 1, September 1, and December 1, 2022, respectively; (i), (j), (k), and (l) represent the global distribution maps of ZWD on March 1, June 1, September 1, and December 1, 2023, respectively;

[0044] Figure 7 : It is a ZWD time series diagram in an embodiment of the present invention, wherein (a) represents the time series diagram of a station with longitude and latitude of (0.5°, 120.5°), (b) represents the time series diagram of a station with longitude and latitude of (0.5°, 300.5°), (c) represents the time series diagram of a station with longitude and latitude of (20.5°, 60.5°), and (d) represents the time series diagram of a station with longitude and latitude of (-20.5°, -60.5°) and (20.5°, 60.5°). DETAILED DESCRIPTION

[0045] In terms of model building methods, it is difficult to characterize the complex changes of ZWD through model functions, while machine learning can process a large amount of complex data and extract the laws of data changes. Previous researchers used machine learning algorithms such as back propagation neural networks and generalized regression neural networks to build superior tropospheric parameter models, proving the superiority of machine learning. Establishing a global long-term forecast model requires processing a large amount of data and preventing overfitting problems. The random forest algorithm has the advantages of high accuracy, strong resistance to overfitting, and the ability to process large amounts of data, which can meet the needs of model building.

[0046] Based on this, the present invention will first use the data processed by Vienna Mapping Functions 3 (VMF3) for three years (2021 to 2023) to draw the ZWD global distribution map and time series map, and conduct a preliminary analysis of the distribution law of ZWD in time and space; then establish a ZWD forecast model based on the random forest algorithm that can predict globally and for a long time; finally, the established forecast model is tested for accuracy.

[0047] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Embodiment 1

[0049] This embodiment discloses a ZWD prediction method based on a random forest algorithm, see Figure 1 ,include:

[0050] S1: Collect tropospheric delay station data worldwide;

[0051] S2: Preprocessing of the collected tropospheric delay station data, including annual accumulation and day conversion and data cleaning;

[0052] S3: Extract the time, tropospheric dry delay and tropospheric wet delay characteristics from the preprocessed tropospheric delay station data;

[0053] S4: Taking time and tropospheric dry delay characteristics as input and tropospheric wet delay characteristics as output, a random forest algorithm model is constructed and trained;

[0054] S5: Use the trained random forest algorithm model to predict ZWD for the area to be tested.

[0055] Specifically, the random forest algorithm and tropospheric wet delay involved in the present invention are described as follows:

[0056] Random forest algorithm: The random forest algorithm adopts the strategy of bootstrapping and aggregation, extracts samples with replacement from the original sample set, constructs multiple different new sample sets, and independently constructs decision trees based on these sample sets to form a random forest containing many decision trees. When a new sample is added, each decision tree in the random forest will independently predict it and give the corresponding prediction results. Finally, the prediction results of multiple decision trees are integrated to obtain the final prediction result of the model. The random forest algorithm is essentially an optimization of the decision tree algorithm. In the random forest algorithm, the decision tree is used as the base learner in the bagging principle, and the training samples of each tree are randomly selected from the overall sample set. This random extraction ensures the diversity of the model in the random forest, and each tree has unique classification and prediction capabilities. Although the samples extracted from each decision tree are different, there is still a certain correlation between them. Therefore, the algorithm further randomly extracts the features selected by each decision tree. This randomness helps to reduce the correlation between decision trees, thereby improving the overall performance of the random forest. The classification ability of a single decision tree is limited, but the random forest contains a large number of randomly generated decision trees. The predicted value or classification result of a new sample can be obtained by summarizing the prediction results of many decision trees, which can significantly improve the prediction ability of the model. Many decision tree algorithms require pruning, but the random forest algorithm uses two random draws, so pruning is not required during the generation of the decision tree, and it will not cause overfitting problems.

[0057] Zenith Wet Delay (ZWD): Zenith Wet Delay (ZWD) refers to the delay caused by the refraction of the signal by water vapor in the atmosphere when the satellite signal passes through the Earth's troposphere. This delay is proportional to the water vapor content in the atmosphere. Therefore, ZWD is an important indicator of the total amount of atmospheric water vapor and is widely used in meteorology, climate research, geophysics, satellite navigation and positioning, etc.

[0058] See also Figure 2 , is a detailed flow chart of the ZWD prediction method based on the random forest algorithm in an embodiment of the present invention. Figure 2 The five main steps involved are described.

[0059] 1. Data Collection

[0060] During the specific implementation process, the data used in this embodiment is the data of the discrete tropospheric delay model-microwave rangefinder-grid-1.0°×1.0° time series site obtained through a professional website. The data includes year, month, day, time point, site longitude and latitude, tropospheric water vapor vertical delay (ah), tropospheric liquid water vertical delay (aw), tropospheric zenith dry delay (Zenith Tropospheric Delay, ZHD) and ZWD information. The time resolution of the data on the website is 6 hours, that is, it includes data at 00:00, 06:00, 12:00 and 18:00 every day. The data source is as follows: Figure 3 As shown (red path).

[0061] Figure 3 In V3GR, a specific version of VMF3 is specifically designed for GNSS systems. It uses ray tracing technology to obtain ZTD. DORIS (Doppler Orbitography and Radiopositioning Integrated by Satellite) is a system for determining satellite orbits. It calculates the speed and position of satellites by measuring the Doppler frequency shift of radio signals emitted by satellites. VLBI (Very Long Baseline Interferometry) is very long baseline interferometry. Multiple observation stations simultaneously track and observe the same target. The observation data is transmitted or recorded on disk in real time, and then transported to the VLBI data processing center for data playback and cross-correlation calculation. The time difference and rate of change of the signal arriving at each observation station can be calculated, and then the angular position of the target can be calculated. SLR (Satellite Laser Ranging) is a technology that determines the distance between the satellite and the ground station by emitting lasers to the satellite and measuring the return time of the reflected light. SLR can provide millimeter-level measurement accuracy and is used to monitor the rotation of the earth, changes in the earth's gravity field, and precise determination of satellite orbits.

[0062] 2. Data Preprocessing

[0063] (1) Conversion between year-to-day and accumulated-year: The original data format is in the format of time point (year, month, day, hour) - station longitude and latitude coordinates, ZWD, and ZHD. Python software is used to preprocess the data, calculate the year-to-day and convert it into the format of station longitude and latitude coordinates - time point (year, month, day, year-to-day and accumulated-year), ZHD, ZWD, and ZTD.

[0064] Before conversion, the data format in each time point file is as shown in Table 1:

[0065] Table 1 Data format before conversion

[0066]

[0067] In Table 1, the first two fields are the grid spatial location information, the third column is the tropospheric water vapor vertical delay, the fourth column is the tropospheric liquid water vertical delay, the fifth column is the tropospheric dry delay, and the last column is the tropospheric wet delay.

[0068] After conversion, the data format in each site file is as shown in Table 2:

[0069] Table 2 Data format after conversion

[0070]

[0071] The first column is the year, the second column is the annual day, the third column is the month, the fourth column is the day, the fifth column is the time, and the last column is the total tropospheric delay.

[0072] (2) Data cleaning: The process of identifying, modifying, replacing or deleting incorrect, incomplete, inaccurate, irrelevant or missing data in a data set. The purpose of data cleaning is to improve data quality and ensure data consistency and accuracy, thereby providing a reliable data foundation for subsequent processing such as data analysis, data mining or machine learning. After inspection, the obtained data has no missing values, outliers, etc.

[0073] 3. Drawing and analysis of the temporal and spatial distribution of tropospheric wet delay

[0074] In this embodiment, the spatiotemporal distribution diagram of tropospheric wet delay is drawn using MATLAB R2021b software, and the spatial distribution diagram of tropospheric wet delay in the world and the time series diagram of selected points are programmed. The distribution of tropospheric wet delay in different seasons, different latitudes, different climate zones, etc. is analyzed. A general understanding of the changes and distribution of tropospheric wet delay and the reasons for this distribution is obtained, which provides a reference for the analysis of experimental results.

[0075] 4. Model Training

[0076] The random forest algorithm is an improvement on the bagging method. The improvement of the random forest algorithm on the bagging algorithm is to introduce additional randomness to enhance the generalization ability of the model and reduce the risk of overfitting. The additional randomness refers to the random selection of features when generating a decision tree. Specifically, the decision tree randomly selects a part of the features for splitting. For each node, the error of the splitting results of different features is compared to select the optimal feature for splitting. Figure 4 shown.

[0077] 5. Accuracy Assessment

[0078] The accuracy indicators such as the sum of squared errors, mean absolute error, mean square error, and root mean square error are calculated to comprehensively evaluate the accuracy of the model.

[0079] The present invention discloses a method for GNSS tropospheric wet delay prediction, constructs a random forest model, and belongs to the field of GNSS comprehensive application. The invention mainly carries out the work of data acquisition, data preprocessing (annual accumulation and day conversion, conversion of raw data into time series data format, etc.), model construction, and accuracy evaluation. Taking tropospheric wet delay as the research object, the time of various stations around the world, tropospheric dry delay, etc. as observation data, and the research method of applying random forest algorithm to tropospheric wet delay regression prediction, a tropospheric wet delay prediction model for global long-term prediction is established.

[0080] The key points of the present invention include two aspects:

[0081] (1) A random forest algorithm model is constructed with time and tropospheric dry delay as features. The present invention uses tropospheric dry delay and time as features, and time is divided into features such as year, month, day, and annual cumulative day. The reason why tropospheric dry delay is used as a feature is that both tropospheric dry delay and tropospheric wet delay are affected by surface temperature and humidity, while tropospheric dry delay is easier to obtain and the measurement data is more accurate. The accuracy of ZHD is used to improve the estimation of ZWD. The random forest algorithm model has advantages in processing large amounts of data on a global scale, and the prediction results are more accurate. This algorithm effectively avoids the problem of overfitting of predictions, and the time used to process large amounts of data is shorter, and the prediction accuracy is higher. In addition, the random forest algorithm can also provide feature importance evaluation, which is extremely important for analyzing the driving factors of ZWD changes.

[0082] (2) The model forecast range is global and the forecast period is up to one month. In the field of ZWD forecast research, meteorological parameter models dominate. However, some regions do not have enough meteorological observation stations or do not have timely updated meteorological data, resulting in low spatial resolution of the model, or under complex terrain and geographical conditions, such models cannot fully consider the influence of terrain factors, resulting in low model prediction accuracy. The accuracy of ZWD long-term prediction is affected by many factors, including weather fluctuations and seasonal changes. In addition, in order to achieve accurate long-term ZWD prediction, it is necessary to rely on accurate and continuous observation data. However, long-term data records around the world are relatively limited, making it difficult to achieve long-term high-temporal resolution and accurate prediction of ZWD. This scheme uses the random forest algorithm to analyze and model three years of global data, and explores its changing laws through machine learning without relying on meteorological data. The random forest algorithm can efficiently process large amounts of data, converge quickly, and have strong prediction timeliness. Its excellent prediction accuracy and ability to handle nonlinear relationships have attracted much attention, and by performing methods such as random sampling, the algorithm can effectively suppress overfitting problems.

[0083] The present invention proposes a tropospheric wet delay prediction model based on a random forest algorithm. Before establishing the model, the data is processed into a time series format, and a global distribution map and a time series map of the tropospheric wet delay are plotted, which intuitively shows the changes in the tropospheric wet delay in different regions and at different times. When establishing the model, tropospheric dry delay, year, month, day, and annual accumulation of days are used as features, tropospheric wet delay is used as the prediction object, and the random forest algorithm is used for training. After the model is established, the test set prediction result graph, the test set regression graph, the test set error histogram, the feature importance graph, and the error curve graph are plotted, and the accuracy indicators such as the sum of squared errors, mean absolute error, mean square error, and root mean square error are calculated to comprehensively judge the accuracy of the model. The random forest algorithm effectively solves the problem of excessive data processing in the global long-term forecast model, while preventing the overfitting problem, and the accuracy assessment shows that the test set prediction result has a high accuracy. The present invention uses observation data from various points around the world within three years to establish a model and predicts the global tropospheric wet delay for the next month, providing a way of thinking and method for a global long-term tropospheric wet delay model.

[0084] The present invention will be further described below through specific examples and accompanying drawings:

[0085] 1. Data Collection

[0086] This implementation will use data from 2021 to 2023 as training samples to predict ZWD data for one month (January 2024), and obtain data from 64,800 sites around the world for three years and one month. The amount of data is relatively large. In the experiment, an EXCEL table is used to list the URLs of ZWD data at different times, and a Shell script is written to download the data through Mobaxterm software.

[0087] 2. Data Preprocessing

[0088] (1) Data processing for drawing global distribution maps:

[0089] Due to the large amount of data, this implementation method selects 12 more representative time nodes when drawing and analyzing the global distribution of ZWD, namely 00:00 on March 1, 2021, 00:00 on June 1, 2021, 00:00 on September 1, 2021, 00:00 on December 1, 2021, 00:00 on March 1, 2022, etc., until 00:00 on December 1, 2023. These time nodes represent the four seasons of spring, summer, autumn and winter from 2021 to 2023 respectively.

[0090] Convert text data into EXCEL table data. Take the table data of the (0.5°, 0.5°) station as an example, as shown in the table.

[0091] Table: 3 Tabular data used to draw the global distribution map of ZWD

[0092]

[0093] (2) Data preprocessing for drawing time series graphs:

[0094] Since the format of the downloaded data is not convenient for drawing the ZWD time series graph and building the random forest algorithm model, Python software was used to convert the time series-site data format into the site-time series data format before the experiment, and the converted data was stored in an EXCEL table. The time series data format in each site file is shown in Table 4.

[0095] Table 4 Tabular data used to draw the ZWD time series graph

[0096]

[0097]

[0098] (2) Feature correlation analysis:

[0099] In addition to the correlation between features and predicted objects, there is also a certain correlation between features. The purpose of analyzing feature correlation is to reduce feature redundancy and check data quality. The feature correlation matrix heat map is drawn as follows: Figure 5 As shown. The correlation between the feature annual cumulative day and month is 1, the correlation between ZHD and the annual cumulative day and month is medium, and the correlation between other features is low. The number of features selected in the implementation method is small, and a small amount of redundancy is accepted to ensure data integrity. The highest correlation between other features is only 0.27, and such feature selection is feasible.

[0100] 3. Drawing and analysis of the temporal and spatial distribution of tropospheric wet delay

[0101] (1) Spatial distribution map

[0102] The global distribution of ZWD at different times is shown in the figure below: Figure 6 As shown in the figure, the darker the blue, the lower the ZWD value, and the darker the orange-red, the higher the ZWD value.

[0103] First, the ZWD values ​​in the equatorial and low-latitude regions are significantly higher than those in the polar and high-latitude regions. Second, the ZWD values ​​in the area where land and ocean meet are higher. Finally, latitude has a greater impact on the ZWD values. Analysis shows that the reasons for this phenomenon are as follows: (1) Water vapor content: Since the equatorial and low-latitude regions are close to water sources such as oceans and tropical rainforests, the evaporation and release of water vapor are more frequent, resulting in a relatively high water vapor content in the equatorial atmosphere, which in turn leads to a higher ZWD. (2) Temperature: The temperature in the equatorial and low-latitude regions is relatively high, and water vapor is easier to evaporate and remain in the atmosphere. The temperature in the polar and high-latitude regions is lower, and the water vapor content in the atmosphere is relatively less. (3) Convective activity: Due to the rotation of the earth and the thermal effects near the equator, atmospheric convection in the equatorial and low-latitude regions is more intense and frequent, resulting in more water vapor being lifted to high altitudes, thereby increasing the tropospheric wet delay. (4) Influence of human factors: Affected by industrial emissions and human activities, the temperature of land and surrounding areas is relatively high, and the ocean around the land provides sufficient water sources. Therefore, water vapor evaporates faster in the area where land and ocean meet, resulting in a higher water vapor content in the atmosphere.

[0104] (2) Time series graph

[0105] In order to more comprehensively reflect the trend of global ZWD changes over time, this implementation method selects some stations to draw time series graphs, namely (0.5°, 0.5°), (0.5°, 60.5°), (0.5°, 120.5°), (0.5°, 180.5°), (0.5°, 240.5°), (0.5°, 300.5°), (20.5°, 0.5°), (20.5°, 40.5°), etc., up to (80.5°, 300.5°). The time series graphs of some points are as follows: Figure 7 As shown, the blue, red and green lines in the figure represent the ZWD data for 2021, 2022 and 2023 respectively.

[0106] Analysis of the rules: (1) In some areas near the equator, the ZWD values ​​from July to November slightly decrease, and are basically stable in other months; while the ZWD values ​​in other areas remain stable throughout the year. It can be seen that due to the influence of latitude factors, the ZWD values ​​near the equator are less affected by seasonal changes. (2) In the northern hemisphere, from January to May, the ZWD value is stable at around 0.1; from June to July, the ZWD value gradually increases; from August to October, the ZWD value remains at a high value, basically between 0.2 and 0.4; from October to November, the ZWD value decreases; from November to December, the ZWD value returns to around 0.1. In the northern hemisphere, June to November is roughly summer and autumn, and the temperature in summer and autumn is higher than that in spring and winter, so the increase in temperature is one of the reasons for the increase in ZWD value. The precipitation in summer is usually higher, and under the influence of high temperature, water vapor evaporates faster than in other seasons, which also causes the ZWD value to increase. (3) In the Southern Hemisphere, the change pattern of ZWD value over time is exactly opposite to that in the Northern Hemisphere. The ZWD value from June to November is smaller than that in other months. The reason for this is that the seasons in the Southern and Northern Hemispheres are opposite. June to November is the winter and spring in the Northern Hemisphere. The temperature is lower, the rainfall is less, and the ZWD value is also lower.

[0107] 4. Model Training

[0108] Take year, annual accumulation day, month, day, time point, and ZHD as input features, and ZWD as output. Calculate the total number of samples, and set the ratio of training set and test set samples to 9:1. Divide the training set and test set. Set the internal parameters of the training model, turn on the prediction based on the out-of-bag error estimate (OOBPrediction = 'on'), calculate the predictor importance based on the out-of-bag error estimate (OOBpredictorImportance = 'on'), select the regression analysis method (Method = 'regression'), set the number of decision trees to 40 (trees = 50), and the minimum number of samples for each leaf node to 5 (leaf = 5). Predict the training set and test set. Calculate the various error parameters of the training set and test set. Output the model prediction results. The image results include the test set regression graph, the test set error histogram, the test set prediction result graph, the feature importance graph, the error curve graph, the training set regression graph, the training set error histogram, and the training set prediction result graph.

[0109] 5. Accuracy Assessment

[0110] This implementation method performs a comprehensive evaluation of multiple error indicators for various sites around the world, as shown in Table 5 for the various indicators of the test set for site (0.5°, 0.5°).

[0111] Table 5 Error indicators of the test set

[0112]

[0113]

[0114] In the formula, n is the total number of samples, y is i is the true value of ZWD, is the predicted value of ZWD.

[0115] The data shows that the model performs relatively well on the test set, with relatively low values ​​for the sum of squared errors, mean squared error, mean absolute error, and root mean square error, indicating that the difference between the model's predicted values ​​and the actual values ​​is small. The average percentage error is only 3.916%, indicating that the model performs well in terms of relative error.

[0116] Embodiment 2

[0117] Based on the same inventive concept, this embodiment discloses a ZWD forecasting device based on a random forest algorithm, comprising:

[0118] Data collection module, used to collect tropospheric delay site data worldwide;

[0119] Data preprocessing module, used to preprocess the collected tropospheric delay site data, including annual accumulation and day conversion and data cleaning;

[0120] A feature extraction module is used to extract time, tropospheric dry delay and tropospheric wet delay features from the preprocessed tropospheric delay station data;

[0121] A model building module is used to build and train a random forest algorithm model by taking time and tropospheric dry delay characteristics as inputs and tropospheric wet delay characteristics as outputs;

[0122] The ZWD forecast module is used to use the trained random forest algorithm model to make ZWD forecasts for the area to be tested.

[0123] Since the device introduced in the second embodiment of the present invention is a device used to implement the ZWD forecasting method based on the random forest algorithm in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, the person skilled in the art can understand the specific structure and deformation of the device, so it is not repeated here. All devices used in the method in the first embodiment of the present invention belong to the scope of protection of the present invention.

[0124] Embodiment 3

[0125] Based on the same inventive concept, the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method described in the first embodiment is implemented.

[0126] Since the computer-readable storage medium introduced in the third embodiment of the present invention is a computer-readable storage medium used to implement the ZWD forecasting method based on the random forest algorithm in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, those skilled in the art can understand the specific structure and deformation of the computer-readable storage medium, so it is not repeated here. All computer-readable storage media used in the method of the first embodiment of the present invention belong to the scope of protection of the present invention.

[0127] Embodiment 4

[0128] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in Embodiment 1 when executing the program.

[0129] Since the computer device introduced in the fourth embodiment of the present invention is a computer device used to implement the ZWD forecasting method based on the random forest algorithm in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, the person skilled in the art can understand the specific structure and deformation of the computer device, so it is not repeated here. All computer devices used in the method of the first embodiment of the present invention belong to the scope of protection of the present invention.

[0130] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0132] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic creative concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A ZWD prediction method based on random forest algorithm, characterized in that: include: Collect tropospheric delay site data worldwide; Preprocess the collected tropospheric delay station data, including annual accumulation and day conversion and data cleaning; Extract the time, tropospheric dry delay and tropospheric wet delay characteristics from the preprocessed tropospheric delay station data; The random forest algorithm model is constructed and trained by taking time and tropospheric dry delay characteristics as input and tropospheric wet delay characteristics as output; The trained random forest algorithm model is used to predict ZWD in the test area.

2. The ZWD prediction method based on random forest algorithm according to claim 1, characterized in that: Preprocess the collected tropospheric delay site data, including: The collected tropospheric delay station data are converted from year to day, and the original data format is converted into the data format of features and targets; Delete incorrect and irrelevant data in the format converted data, and complete incomplete data.

3. The ZWD prediction method based on random forest algorithm according to claim 1, characterized in that: The preprocessed tropospheric delay site data includes time series-site data and site-time series data. After preprocessing the collected tropospheric delay site data, the method further includes: Draw a spatial distribution map based on time series-site data and analyze the spatial distribution pattern; Draw a time series graph based on the site-time series data and analyze the temporal variation trend.

4. The ZWD prediction method based on random forest algorithm according to claim 1, characterized in that: The time, tropospheric dry delay and tropospheric wet delay characteristics are extracted from the preprocessed tropospheric delay station data, including: Extract time data from the preprocessed tropospheric delay station data and divide the time into months, days, annual cumulative days and moments; Tropospheric dry delay and tropospheric wet delay are extracted from the preprocessed tropospheric delay station data.

5. The ZWD prediction method based on random forest algorithm according to claim 1, characterized in that: Before taking the time and tropospheric dry delay features as input and the tropospheric wet delay features as output to construct and train the random forest algorithm model, the method also includes: performing feature correlation analysis on the extracted time and tropospheric dry delay, and selecting features that meet preset conditions as input.

6. The ZWD prediction method based on random forest algorithm according to claim 1, characterized in that: The time and tropospheric dry delay characteristics are used as input, and the tropospheric wet delay characteristics are used as output to build and train the random forest algorithm model, including: Construct the original sample set based on the selected features; Samples are extracted from the original sample set with replacement to construct multiple different new sample sets. Decision trees are independently constructed based on these new sample sets to form a random forest containing many decision trees. When new samples are added, each decision tree in the random forest independently predicts them and gives corresponding prediction results. Finally, the prediction results of multiple decision trees are integrated to obtain the final prediction result of the model.

7. The ZWD prediction method based on random forest algorithm according to claim 1, characterized in that: The method further comprises: The accuracy of the constructed random forest algorithm model was tested.

8. A ZWD forecasting device based on random forest algorithm, characterized in that: include: Data collection module, used to collect tropospheric delay site data worldwide; Data preprocessing module, used to preprocess the collected tropospheric delay site data, including annual accumulation and day conversion and data cleaning; A feature extraction module is used to extract time, tropospheric dry delay and tropospheric wet delay features from the preprocessed tropospheric delay station data; A model building module is used to build and train a random forest algorithm model by taking time and tropospheric dry delay characteristics as inputs and tropospheric wet delay characteristics as outputs; The ZWD forecast module is used to use the trained random forest algorithm model to make ZWD forecasts for the area to be tested.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the ZWD prediction method based on the random forest algorithm as claimed in any one of claims 1 to 7 is implemented.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the ZWD prediction method based on the random forest algorithm as claimed in any one of claims 1 to 7 is implemented.

Images