Seabed suspension concentration prediction method based on random forest

Abstract

The invention discloses a seabed suspension concentration prediction method based on a random forest. The method comprises the following steps: carrying out long-time-sequence SSC in-situ observation on a target sea area; collecting tide level, wave, ocean current and wind field data during in-situ observation; establishing a correlation model between the SSC and the wind field, the tide level, the wave and the ocean current by using a random forest model; forecasting a wind field in the next 24 hours by using a WRF meteorological model; utilizing a Matlab toolkit to forecast the tide level; forecasting waves through the SWAN model by using the forecasted wind field data; forecasting the ocean current through an FVCOM model by utilizing the forecast tide level and wind field result; and predicting the seabed SSC of the target sea area in the next 24 hours based on the trained random forest model and wind field, tide level, wave and ocean current forecast results. According to the invention, original large-area and long-time field observation is replaced, and basic data can be provided for forecasting and early warning of scouring and silting of offshore port channels, stability of subsea pipelines, coastal water quality environment and coast water and soil loss.

Description

Technical Field

[0001] The invention relates to a method for predicting seafloor suspended body concentration based on random forest, and belongs to the technical field of marine geological exploration and water resources monitoring. Background Art

[0002] The concentration of suspended matter in the ocean is an important parameter for studying the stability and safety of the seabed. The concentration of suspended matter can provide important indicative significance for the scouring and silting of offshore port channels, the stability of submarine pipelines, coastal water quality environment, coastal soil erosion, etc. The suspended matter concentration (SSC) seriously affects the evolution of scouring and silting in estuary coastal zones. Therefore, the study of SSC is of great significance for the design of offshore projects (submarine pipelines, offshore oil and gas platforms) and the dredging of port channels.

[0003] At present, the measurement of seabed SSC mainly adopts two methods: in-situ observation by a bottom-mounted system and returning to the room for measurement after sampling. In-situ observation by a bottom-mounted system refers to the real-time observation of SSC by fixing optical instruments such as turbidity meters on a bottom-mounted observation platform. After sampling, returning to the room for measurement first requires manual use of traditional hydrological winches to obtain bottom water samples of the target sea area every hour, and then returning to the room for filtration, drying, and measurement. There are great problems in the implementation of the above two methods on site. Method 1 is difficult to ensure the safety of instruments and equipment for long-term observation. Extreme weather events will cause changes such as sediment liquefaction and resuspension on the seabed, which will greatly affect the stability of the bottom-mounted platform, making it difficult to ensure instrument safety and data quality; Method 2 requires a lot of manpower and material resources, and it is difficult to observe at a fixed point for a long time. Therefore, the existing technology is difficult to sustain both economically and in terms of time.

[0004] At present, the inversion and prediction of SSC through big data and machine learning will greatly improve efficiency and reduce costs. However, the existing machine learning prediction of water body parameters is mostly concentrated in the field of water quality. For example, China Power Construction Group East China Survey and Design Institute Co., Ltd. designed a "water quality forecasting method for estuaries and coasts based on hydrodynamic water quality coupling model" (CN119227868A). This method mainly extracts indicators through the simulation results of the FVCOM-ERSEM model to obtain the spatiotemporal variation law of pollutants in the estuary and coastal areas under different factors and conditions within a set time, which does not include SSC. "A reservoir water quality concentration budget method based on weather forecast" (CN202311834545.8) designed by Wuhan Data Intelligence Research Institute is more inclined to mathematical statistics and lacks a mechanism model for mechanism identification. "A prediction method and system for the maximum turbidity zone" (CN202110625053.2) designed by Sun Yat-sen University and Guangdong Provincial Laboratory of Southern Marine Science and Engineering (Zhuhai) mainly determines the maximum turbidity zone through remote sensing and measured data, and cannot make a fine forecast of the single-point bottom SSC.

[0005] In summary, it is extremely necessary to develop a new method for predicting seabed SSC data, which can not only reduce the original observation cost, but also predict SSC for a period of time in the future and provide technical support for marine engineering construction. Summary of the invention

[0006] In order to overcome the defects of difficulty and high cost of in-situ observation in the ocean, the present invention provides a method for predicting seafloor suspended matter concentration based on random forest. This invention mainly trains a random forest model through measured SSC and hydrodynamic data, and combines it with well-verified forecast data to predict SSC in the next 24 hours.

[0007] The present invention comprises the following steps: 1) Conduct long-term in-situ observations of seafloor suspended matter concentration (SSC) in the target sea area; 2) Collect the measured tide level, wave, current and wind field (wind speed and direction) data from the public dataset (CFSR) during the SSC observation period in the target sea area; 3) Based on the above data, a random forest model is used to establish a correlation model between SSC and the data obtained in step 2; 4) Use Matlab T_tide toolkit to predict the tide level in the target sea area; 5) Forecast the wind field in the next 24 hours using the Weather Research and Forecasting (WRF) meteorological model; 6) Using the predicted wind field data, the waves in the target sea area are predicted through the SWAN model; 7) Using the forecast tide and wind field results, the FVCOM model is used to predict the ocean currents in the target sea area; 8) Predict the seabed SSC of the target sea area within the next 24 hours based on the trained random forest model and forecast results (wind field, tide level, waves, and currents).

[0008] In step (1), the long-term in-situ observation of seafloor suspended matter concentration SSC is a continuous observation of not less than 2 months, by placing a tripod equipped with an optical backscatter turbidity meter OBS on the seafloor to measure SSC, and converting the optical signal into SSC by data fitting.

[0009] The step (2) is to place an acoustic Doppler current meter in a tripod at the bottom of the base while performing in-situ observation in step (1) to collect tide, wave and current data of the station; and to extract wind field data consistent with the SSC observation time by combining the wind field data of the CFSR data set.

[0010] In step (3), the ratio of the training set to the test set is 8:2, and the data set is randomly rearranged to ensure the randomness of the data.

[0011] The WRF meteorological model described in step (5) includes 1) a meteorological data acquisition module, which acquires global forecast field data and observation data; 2) a meteorological data preprocessing module, which processes and obtains model background field data and boundary data; 3) an atmospheric environment parameter numerical prediction module, which completes the forecast of atmospheric environment parameters for the next 7 days; and 4) a numerical prediction post-processing module, which completes the processing and screening of the original data and outputs the product data of the required variables.

[0012] The output variables of the WRF meteorological model include: 10m wind speed components u, v -- U10, V10.

[0013] In step (6), when the SWAN model is running, the target sea area is horizontally divided using unstructured triangular grids, and grid encryption is performed near the coast so that the protruding coastline can be finely portrayed, and the SWAN model uses the wind field predicted in step (5) as the driving force for simulation.

[0014] In step (7), the scope and settings of the FVCOM model are the same as those of step (6), and the FVCOM model uses the wind field predicted in step (5) and the tide level predicted in step (4) as driving forces for simulation.

[0015] The innovation and advantages of the present invention are: 1) A random forest-based method for predicting seafloor suspended matter concentration is proposed, which can use machine learning to predict suspended sediment concentration, replacing the original large-scale, long-term field observations and reducing costs. The machine learning method is coupled with weather forecasts to achieve the above goals well.

[0016] 2) Based on the verified accurate forecast meteorological data, the hydrodynamic data is calculated to obtain the suspended matter concentration data, which makes up for the problem that the original method can only analyze the existing suspended matter concentration data and cannot make future predictions. This prediction can provide basic data for the scouring and silting of offshore port channels, the stability of submarine pipelines, coastal water quality environment, and the forecast and warning of coastal soil erosion.

[0017] 3) The data reserve required for prediction can be completed based on the measured data once. Compared with the field observation method, the present invention does not require multiple manual observations, which is simpler and more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the present invention.

[0019] Figure 21 is the conversion of OBS optical parameters to SSC shown in the embodiment, where (a) is the linear regression of laboratory optical backscatter sensor (OBS) turbidity and suspended sediment concentration (SSC); (b) is the linear regression of Doppler velocimetry and signal-to-noise ratio (SNR) recorded by SSCOBS.

[0020] Figure 3 This is a diagram of the validation effect of the random forest model, where (a) is the validation of the training set training results; (b) is the validation of the prediction set prediction results.

[0021] Figure 4 It is a comparison between the tide level predicted by Matlab T_tide toolkit and the measured tide level.

[0022] Figure 5 This is the WRF atmospheric model forecast flow chart.

[0023] Figure 6 The spatial correlation coefficient between the WRF model forecast results and the public data set changes with the forecast duration (the blue line is the east component of the wind field; the yellow line is the north component of the wind field; the green line is the 2-meter temperature field; the red line is the 500hPa potential field).

[0024] Figure 7 It is the comparison between the SWAN wave mode simulation results and the measured data.

[0025] Figure 8 This is the comparison between the FVCOM ocean current simulation results and the measured data, where (a) is the measured ocean current u component; (b) is the measured ocean current v component; (c) is the simulated ocean current u component; (d) is the simulated ocean current v component.

[0026] Fig. 9 is the SSC prediction result, where (a) is the comparison result between the measured and predicted SSC; (b) is the difference between the measured and predicted SSC. DETAILED DESCRIPTION

[0027] The following is a specific implementation description of the present invention in combination with the content of the invention, the accompanying drawings and the specific implementation methods. It should be emphasized that the following specific implementation description is only exemplary and does not limit the scope of use and application of the present invention. Figure 1 As shown in the flow chart, the present invention includes the following steps.

[0028] 1) Conduct long-term in-situ observations of seafloor suspended matter concentration (SSC) in the target sea area; Taking the relevant marine environment near the Yellow River Delta as an example, in the previous study, a three-month bottom-based in-situ observation was conducted in the waters near the mouth of the Yellow River, and an SSC in-situ observation data set with a time resolution of 1 hour was obtained. The observation time was from January 14, 2020 to April 19, 2020. A bottom-based tripod equipped with an optical backscatter sensor (OBS) was deployed at a distance of 80 cm from the bottom to measure SSC and record data in real time. The optical signal was converted into SSC using the following formula (such as Figure 2 shown): Y=1.7674x-232.1597 Where x is the optical signal and y is the SSC.

[0029] 2) Collect tide, wave, current and wind field (wind speed and direction) data from a public dataset (CFSR) during the SSC in situ observation period in the target sea area; While observing SSC in (1), an acoustic Doppler current meter (Nortek 600 kHz Acoustic Wave and Current, AWAC, also known as "Langlong") was deployed in the tripod at the bottom of the station to collect tide, wave, and current data at the station; combined with the wind field data of the CFSR dataset, the wind field data consistent with the SSC observation time were extracted. The time resolution of the above data was also selected as 1 hour.

[0030] 3) Based on the above data, a random forest model is used to establish a correlation model between SSC and the data obtained in step 2; The training set of the random forest model is 1744 groups of data, and the test set is 436 groups of data sets, with a ratio of 8:2. To ensure the randomness of the data, the randperm function in Matlab is used to randomly rearrange the data sets. The model verification results after training are good, and the correlation can reach more than 0.8. The effect is as follows Figure 3 shown.

[0031] 4) Use Matlab T_tide toolkit to predict the tidal level of the target sea area. The prediction effect is as follows: Figure 4 As shown; 5) Use the Weather Research and Forecasting (WRF) meteorological model to forecast weather data for the next 24 hours; The atmospheric forecast system for the Yellow Sea and Bohai Sea regions was constructed using the WRF regional atmospheric model. The forecast area is 115ºE~132ºE, 28ºN~42ºN. The model has a horizontal resolution of 10 km, a total number of horizontal grid points of 155×144, and 50 vertical layers. Terrain-following coordinates are used, the pressure at the top of the model is top = 5000Pa, and the projection method is Lambert projection. The initial field and boundary of the model are derived from the Global Forecast System-Final Operational GlobalAnalysis (GFS-FNL) data.

[0032] The WRF mode operation process is shown in Figure 5 , mainly including: 1. Meteorological data acquisition module, which obtains global forecast field data and observation data; 2. Meteorological data preprocessing module processes and obtains model background field data and boundary data; 3. Atmospheric environment parameter numerical prediction module completes the forecast of atmospheric environment parameters for the next 7 days; 4. Numerical forecast post-processing module completes the processing and screening of original data and outputs the product data of required variables (Table 1).

[0033] Table 1 WRF model output variables Variable Name Variable Description unit time Timestamp string XLAT、XLON The latitude and longitude of the grid center point PSFC Surface pressure Pa U10, V10 10m wind speed components u, v m / s TSK Surface temperature K GLW Downward longwave flux from the ground <![CDATA[W / m 2 ]]> SWDOWN Shortwave flux downward from the ground <![CDATA[W / m 2 ]]> TOTAL Total precipitation mm SFCEVP Cumulative surface evaporation <![CDATA[kg / m 2 ]]> RH Relative humidity % .

[0034] The prediction effect of the WRF model wind field was verified using a public wind field dataset ( Figure 6 ). Although the forecast effect of the surface wind field (U10 and V10) shows a downward trend overall, the correlation coefficient of the surface wind field forecast within 1 day is relatively high, reaching above 0.8, which can meet the subsequent forecast work.

[0035] 6) Using the predicted wind field data, the waves in the target sea area are predicted through the SWAN model; The calculation range of the model in this paper is 117.5°~122.5°E, 37°~41°N, including the entire Bohai Sea area, and the open boundary is selected near the Dalian-Yantai line in the Bohai Strait. This paper uses unstructured triangular grids to divide the calculation area horizontally, and performs grid encryption near the coast, and protruding coastlines such as ding dams are also finely portrayed. The entire area contains a total of 83,601 triangular grids and 45,455 grid nodes. The horizontal resolution is up to 80m near the seawall of the Yellow River Delta and 5000m at the open boundary in the open sea. The vertical direction is divided into 7 equally spaced σ layers. During the forecast process, the model uses the wind field predicted in step 5 as the driving force for simulation. The wave forecast effect is as follows Figure 7 As shown; 7) Using the forecast tide and wind field results, the FVCOM model is used to predict the ocean currents in the target sea area; The model scope and settings are the same as step 6. During the forecast process, the model uses the wind field predicted in step 5 and the tide level predicted in step 4 as the driving force for simulation. Figure 8 As shown; 8) Based on the trained random forest model and forecast results (wind field, tide level, waves, and ocean currents), predict the seabed SSC of the target sea area within the next 24 hours. Fig. 9 shown.

[0036] In the observation area in (1), the same observation method was used to observe the bottom SSC data on November 15, 2021 to verify the accuracy of the random forest prediction results. Based on this method, the accuracy of predicting the seabed SSC in the target sea area within the next 24 hours can reach 0.79 ( Fig. 9 ), which can effectively reflect the SSC change trend of the target sea area and greatly reduce the cost of on-site observation.

Claims

1. A method for predicting seafloor suspended matter concentration based on random forest, characterized by The following steps are involved: (1) Conduct in-situ observations of long-term seafloor suspended matter concentrations in the target sea area; (2) Collect tide, wave, current and wind data from the public dataset CFSR during the SSC in situ observation period in the target sea area; (3) Based on the above data, a random forest model was used to establish a correlation model between SSC and wind field, tide level, waves, and ocean currents; (4) Use Matlab T_tide toolkit to predict the tidal level of the target sea area; (5) Use the WRF meteorological model to predict the wind field in the next 24 hours; (6) Using the predicted wind field data, the SWAN model is used to predict the waves in the target sea area; (7) Using the predicted tide and wind field results, the FVCOM model is used to predict the ocean currents in the target sea area; (8) Predict the seabed SSC of the target sea area within the next 24 hours based on the trained random forest model and the wind field, tide level, wave, and current forecast results.

2. The prediction method according to claim 1, characterized in that In step (1), the long-term in-situ observation of seafloor suspended matter concentration SSC is a continuous observation of not less than 2 months, by placing a tripod equipped with an optical backscatter turbidity meter OBS on the seafloor to measure SSC, and converting the optical signal into SSC by data fitting.

3. The prediction method according to claim 1, characterized in that The step (2) is to place an acoustic Doppler current meter in a tripod at the bottom of the base while performing in-situ observation in step (1) to collect tide, wave and current data of the station; and to extract wind field data consistent with the SSC observation time by combining the wind field data of the CFSR data set.

4. The prediction method according to claim 1, characterized in that In step (3), the ratio of the training set to the test set is 8:2, and the data set is randomly rearranged to ensure the randomness of the data.

5. The prediction method according to claim 1, characterized in that The WRF meteorological model described in step (5) includes 1) a meteorological data acquisition module, which acquires global forecast field data and observation data; 2) a meteorological data preprocessing module, which processes and obtains model background field data and boundary data; 3) an atmospheric environment parameter numerical prediction module, which completes the forecast of atmospheric environment parameters for the next 7 days; and 4) a numerical prediction post-processing module, which completes the processing and screening of the original data and outputs the product data of the required variables.

6. The prediction method according to claim 5, characterized in that The output variables of the WRF meteorological model include: 10m wind speed components u, v - U10, V10.

7. The prediction method according to claim 1, characterized in that In step (6), when the SWAN model is running, the target sea area is horizontally divided using unstructured triangular grids, and grid encryption is performed near the coast so that the protruding coastline can be finely portrayed, and the SWAN model uses the wind field predicted in step (5) as the driving force for simulation.

8. The prediction method according to claim 1, characterized in that In step (7), the scope and settings of the FVCOM model are the same as those of step (6), and the FVCOM model uses the wind field predicted in step (5) and the tide level predicted in step (4) as driving forces for simulation.

Images