A numerical forecast correction method based on ship-based cruise observation

Abstract

The present invention relates to a numerical forecast correction method based on ship-based cruise observation, comprising: 1. selecting a ship-based cruise area, and using WRF to output WRF numerical simulation results; 2. calculating the total wind speed uv, and using a sliding average algorithm to process the second-by-second measured data into t-second interval data; 3. performing spatiotemporal matching between the t-second interval data processed in step 2 and the WRF numerical simulation results, and constructing a data set; 4. constructing a random forest model, and training the random forest model with a training data set in the data set to obtain a numerical forecast correction model; 5. inputting a test data set in the data set into the numerical forecast correction model, and performing an accuracy test on the numerical forecast correction model, and then using it for numerical forecast correction of ship-based cruise observation. The present invention uses a machine learning method to correct WRF numerical simulation wind speed data based on ship-based cruise observation data, so that the numerical simulation data is closer to the true value, and accurate correction of the offshore numerical simulation wind field is achieved.

Description

Technical Field

[0001] The present invention relates to the field of ship navigation wind field forecasting, and in particular to a numerical forecast correction method based on ship-based cruise observation. Background Art

[0002] The sea surface wind field is an important reference factor for formulating ship route plans and ensuring navigation safety. Strong winds at sea often cause meteorological disasters such as disastrous waves and even storm surges, causing accidents such as ship capsizing. Therefore, accurate forecasts of offshore wind fields are of great significance to ensuring the activities of ships at sea and offshore engineering.

[0003] Numerical model forecast is one of the important means of offshore wind field forecast. However, due to the single means of measuring the offshore wind field, most of the measured data on the sea surface are obtained by ships and ocean buoys, resulting in a serious lack of measured data on the sea surface and high altitude, which makes it very difficult for the numerical model to construct a true initial field. In addition, factors such as the assimilation of multi-source observation data, computing resources, and the level of forecasters also limit the accuracy of numerical model forecasts, resulting in a certain error between the forecast value and the actual value. Therefore, it is of great significance to study effective and scientific numerical forecast correction methods to improve the accuracy of meteorological forecasts for the safety of marine activities.

[0004] Research has been conducted at home and abroad on correction technology. Generally, forecast errors can be divided into systematic errors and non-systematic errors:

[0005] Systematic errors do not change with time and are generally obtained by averaging a large number of forecast errors over time, representing the drift of the equilibrium state of the numerical model relative to the actual climate state.

[0006] The non-systematic error is the part that varies with the flow pattern, depends on the atmospheric state variables, and contains random errors.

[0007] The commonly used method for correcting systematic errors is to calculate the average value of a large number of return errors (considering seasonal and daily changes) and superimpose it on the current output result of the corresponding forecast time. However, this method does not consider the nonlinear interaction between internal and external errors in the model integration process. The correction of non-systematic errors considers the change law of errors over time. It is believed that in the short term, the statistical characteristics and spatiotemporal distribution characteristics of the errors do not change much, and recent data and model results are used for correction. However, due to the difficulty in obtaining marine data, it is difficult to obtain long-term measured data that matches the simulation data, and the error changes nonlinearly over time, resulting in unstable correction effects of traditional methods. Summary of the invention

[0008] In order to improve the accuracy of numerical forecasts, the present invention provides a numerical forecast correction method based on ship-based cruise observations. Based on the ship-based cruise measured data and the nonlinear fitting ability of machine learning, the data characteristics of historical measured data and forecast data are mined to correct the simulated forecast wind field of the mesoscale numerical weather forecast model WRF to improve the accuracy of numerical forecasts.

[0009] The technical objectives of the present invention are achieved through the following technical solutions:

[0010] A numerical forecast correction method based on ship-based cruise observation, the method comprising:

[0011] Step 1: Select a ship-based navigation area, use WRF to perform numerical simulation of the offshore wind field in the ship-based navigation area, and output the WRF numerical simulation results at a time interval of t seconds;

[0012] Step 2: Calculate the total wind speed uv based on the measured wind speed data of the ship base.

[0013] The sliding average algorithm is used to calculate the uv and forecast influencing factor data for t seconds of sliding average, and the second-by-second measured data is processed into t-second interval data;

[0014] Step 3, perform temporal and spatial matching of the t-second interval data processed in step 2 with the WRF numerical simulation results output in step 1 to construct a data set;

[0015] Step 4: construct a random forest model, divide the constructed data set into a training data set and then a test data set, and use the training data set to train the random forest model to obtain a numerical forecast correction model;

[0016] Step 5: Input the test data set constructed in step 3 into the numerical forecast correction model constructed in step 4 to perform an accuracy test on the numerical forecast correction model. If the accuracy test meets the requirements, the numerical forecast correction model is used for the numerical forecast correction of ship-based cruise observations.

[0017] Furthermore, in step 2, abnormal data is pre-eliminated when calculating the total wind speed uv.

[0018] Furthermore, when eliminating abnormal data, integrity checks and internal consistency checks are performed on the ship-based measured wind speed data.

[0019] Furthermore, in step 2,

[0020] u=-wspd*sin(wdir),

[0021] v = -wspd*cos(wdir),

[0022]

[0023] Among them, u is the radial wind, v is the zonal wind, wspd is the measured wind speed, wdir is the measured wind direction, and uv is the total wind speed.

[0024] Furthermore, in step 3, when performing time-space matching, the longitude, latitude and observation time of the ship-based measurement point are used as a reference to extract the data of the point closest to the observation point at the same time in the WRF numerical simulation results.

[0025] Further, in step 2, the forecast influencing factor data include 2-meter temperature, 2-meter relative humidity, 10-meter radial wind u, 10-meter zonal wind v, sea level pressure, and temperature, relative humidity, U, V and altitude of 925hPa, 850hPa, 700hPa and 500hPa.

[0026] Furthermore, the dataset is represented as a two-dimensional matrix:

[0027] Z m×n ={(x 1,1 ,x 1,2 ,…,x 1,n-1 ,y 1 ),(x 2,1 ,x 2,2 ,…,x 2,n-1 ,y 2 ),…(x i,1 ,x i,2 ,…,x i,n-1 ,y i ),(x m,1 ,x m,2 ,…,x m,n -1,y m )}, where (x i,1 ,x i,2 ,…,x i,n-1 ,y i ) is the i-th data extracted from the WRF simulation results, i ranges from 1 to m, m is the total number of data, n is the number of characteristic variables, y i is the measured total wind speed value corresponding to the i-th data.

[0028] Furthermore, in step 4, the constructed data set is divided into a training data set and a test data set in a ratio of 4:1.

[0029] Furthermore, in step 4, building a random forest model includes:

[0030] Step 4.1, randomly select N training data sets from the training data set by using the bootstrap resampling method, and the N training data sets have M feature attributes;

[0031] Step 4.2, randomly select a training data set to build, and randomly select m feature attributes from M feature attributes, based on the decision tree model built by the selected training data set, select an optimal feature attribute from the m feature attributes as a partition node, and generate a new decision tree model child node;

[0032] Step 4.3, repeat step 4.2 until the decision tree model is generated, and build several decision tree models according to steps 4.1-4.3;

[0033] Step 4.4: Input the training data into all decision tree models, calculate the mean of the prediction values ​​of all decision tree models, and obtain the final prediction result of the random forest model.

[0034] Furthermore, in step 5, the mean percentage error MAPE and the coefficient of determination R 2 Evaluate the accuracy of the numerical forecast correction model:

[0035]

[0036]

[0037] Among them, Y i is the measured total wind speed value of the ith measured point, is the predicted total wind speed value of the ith measured point, is the average of n true values; the smaller the MAPE, the better the model fitting effect; R 2 It can evaluate the matching degree between the predicted value and the true value, R 2 The closer to 1, the better the model effect.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The present invention is based on ship-based cruise observation data and utilizes the nonlinear fitting ability of machine learning methods to construct a second-level data set and correct the WRF numerical simulation wind speed, making the numerical simulation data closer to the true value. This overcomes the shortcomings of insufficient marine observation information and insufficient accuracy of traditional correction methods, and realizes accurate correction of offshore numerical simulation wind fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of the data forecast correction method based on ship-based cruise observation of the present invention.

[0041] Figure 2 It is a schematic diagram of the range of numerical model simulation in the present invention.

[0042] Figure 3 Schematic diagram of the measured data trajectory in the embodiment of the present invention.

[0043] Figure 4 It is a structural schematic diagram of the random forest model in the present invention. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is further described below in conjunction with specific embodiments:

[0045] A numerical forecast correction method based on ship-based cruise observation, such as Figure 1 As shown, the method includes:

[0046] Step 1: Select the ship-based navigation area, use WRF to perform numerical simulation of the offshore wind field in the ship-based navigation area, and output the WRF numerical simulation results at t second time intervals; WRF (Weather Research and Forecasting, referred to as WRF, is a widely used numerical weather prediction and atmospheric research tool developed by the National Center for Atmospheric Research, the National Oceanic and Atmospheric Administration and its partners. It is used to simulate short-term weather forecasts, atmospheric processes and long-term climate simulations. WRF is more accurate than traditional numerical weather prediction models and has higher temporal and spatial resolution. It is an important tool for the meteorological and atmospheric research community.)

[0047] The details are as follows:

[0048] Obtain the static geographic data required for WRF numerical simulation. The numerical simulation area is as follows: Figure 2 As shown in the figure, with 38.616°N, 121.916°E as the central longitude and latitude, the three-layer nesting technology is used to divide the numerical simulation area grid, the outermost grid number (resolution) is 150×150 (9km), the second outermost d02 grid number (resolution) is 199×199 (3km), and the innermost d03 grid number (resolution) is 169×169 (1km). The numerical simulation strategy is determined to start at 06UTC, the model is 30 hours, the first 6 hours are discarded as spin-up, the model vertical level is set to 45 layers, the outermost time step of the model is 54s, and the simulation time is determined to be from 14:40 on August 1, 2021 (Beijing time) to 11:30 on August 2, 2021 (Beijing time), and the WRF numerical simulation results of the ship-based sailing area are output at 10s time intervals.

[0049] Step 2: Calculate the total wind speed uv based on the ship-based measured wind speed data. Preferably, abnormal data is pre-eliminated when calculating the total wind speed uv. When eliminating abnormal data, an integrity check and an internal consistency check are performed on the ship-based measured wind speed data.

[0050] Integrity check: For abnormal wind speed and wind direction data caused by anemometer machine failure and other possible factors, the abnormal wind speed and wind direction data need to be eliminated. For example, the data needs to be eliminated if it is blank, missing, or out of the observation time range.

[0051] Internal consistency check: There are three situations: one is wind direction > 360° or wind direction < 0°, one is wind speed is zero and wind direction ≥ 0°, and the other is wind speed > 75m / s. Data with any of the above three situations must be eliminated.

[0052] The schematic diagram of the measured data trajectory in this embodiment is as follows Figure 3 As shown, the trajectories are concentrated in the area centered at 38.6°N, 121.9°E. After eliminating abnormal data, 82,866 valid measured data are obtained.

[0053] Then use the sliding average algorithm to calculate the UV and forecast influencing factor data for t-second sliding average, and process the second-by-second measured data into t-second interval data; the details are as follows:

[0054] u=-wspd*sin(wdir),

[0055] v = -wspd*cos(wdir),

[0056]

[0057] Among them, u is the radial wind, v is the zonal wind, wspd is the measured wind speed, wdir is the measured wind direction, and uv is the total wind speed.

[0058] Using the sliding average algorithm, the calculated total wind speed uv and forecast influencing factor data are subjected to a 10-second sliding average. The forecast influencing factor data involved include: 2-meter temperature, 2-meter relative humidity, 10-meter radial wind u, 10-meter zonal wind v, sea level pressure, and 925hPa temperature, 925hPa relative humidity, 925hPa radial wind u, 925hPa zonal wind v, 925hPa altitude, 850hPa temperature, 850hPa relative humidity, 850hPa radial wind u, 850hPa zonal wind v, 850hPa altitude, 700hPa temperature, 700hPa relative humidity, 700hPa radial wind u, 700hPa zonal wind v, 700hPa altitude, 500hPa temperature, 500hPa relative humidity, 500hPa radial wind u, 500hPa zonal wind v, 500hPa altitude.

[0059] Step 3: Process the measured data second by second into 10-second interval data and match them with the WRF numerical simulation results of the ship-based cruise area output at 10-second intervals to construct a data set.

[0060] When performing space-time matching:

[0061] The WRF numerical simulation results are multivariate regional data with an interval of 10 seconds, covering the observation route area. Based on the longitude, latitude and observation time of the ship-based measurement point, the data of the point closest to the observation point at the same time in the WRF numerical simulation results are extracted according to the minimum distance criterion. The main variables extracted include: 2-meter temperature, 2-meter relative humidity, 10-meter radial wind u, 10-meter zonal wind v, sea level pressure, and 925hPa temperature, 925hPa relative humidity, 925hPa radial wind u, 925hPa zonal wind v, 925hPa height, 850h Temperature of 800hPa, relative humidity of 850hPa, radial wind u of 850hPa, latitudinal wind v of 850hPa, altitude of 850hPa, temperature of 700hPa, relative humidity of 700hPa, radial wind u of 700hPa, latitudinal wind v of 700hPa, altitude of 700hPa, temperature of 500hPa, relative humidity of 500hPa, radial wind u of 500hPa, latitudinal wind v of 500hPa, altitude of 500hPa.

[0062] The dataset construction process is as follows:

[0063] Z m×n ={(x 1,1 ,x 1,2 ,…,x 1,n-1 ,y 1 ),(x 2,1 ,x 2,2 ,…,x 2,n-1 ,y 2 ),…(x i,1 ,x i,2 ,…,x i,n -1,y i ),(x m,1 ,x m,2 ,…,x m,n-1 ,y m )}, where (x i,1 ,x i,2 ,…,x i,n-1 ,y i ) is the i-th data extracted from the WRF simulation results, i ranges from 1 to m, m is the total number of data, n is the number of characteristic variables, y iis the measured total wind speed value corresponding to the i-th data. The characteristic variables include: 2-meter temperature, 2-meter relative humidity, 10-meter radial wind u, 10-meter zonal wind v, sea level pressure, and 925hPa temperature, 925hPa relative humidity, 925hPa radial wind u, 925hPa zonal wind v, 925hPa height, 850hPa temperature, 850hPa relative humidity, 850hPa radial wind u, 850hPa zonal wind v, 850hPa height, 700hPa temperature, 700hPa relative humidity, 700hPa radial wind u, 700hPa zonal wind v, 700hPa height, 500hPa temperature, 500hPa relative humidity, 500hPa radial wind u, 500hPa zonal wind v, 500hPa height.

[0064] Step 4: Construct a random forest model. Randomly divide the constructed data set into a training data set and a test data set according to a ratio of 4:1. Use the training data set to train the random forest model to obtain a numerical forecast correction model. The structure of the random forest model is as follows: Figure 4 As shown in the figure, the random forest model consists of many decision tree models. After each data set is input, each decision tree model makes judgments and classifications and obtains a result. The one with the most classifications among the output results of the decision tree model is the final output result. The training process of the random forest model is as follows:

[0065] Step 4.1, randomly select N training data sets from the training data set by using the bootstrap resampling method, and the N training data sets have M feature attributes;

[0066] Step 4.2, randomly select a training data set to build, and randomly select m feature attributes from M feature attributes, based on the decision tree model built by the selected training data set, select an optimal feature attribute from the m feature attributes as a partition node, and generate a new decision tree model child node;

[0067] Step 4.3, repeat step 4.2 until the decision tree model is generated, and establish several decision tree models according to steps 4.1-4.3; by setting the decision tree height of the decision tree model, or setting the number of samples contained in each node to reach the set threshold, the generation of the current decision tree model is completed after the set conditions are met.

[0068] Step 4.4: Input the training set data into the random forest model, calculate the mean of all decision tree model predictions, and obtain the final prediction result of the random forest model.

[0069] Step 5: Input the test data set constructed in step 3 into the numerical forecast correction model constructed in step 4 to perform an accuracy test on the numerical forecast correction model. If the accuracy test meets the requirements, the numerical forecast correction model is used for the numerical forecast correction of ship-based cruise observations.

[0070] Specifically, the mean percentage error (MAPE) and the coefficient of determination (R) 2 Evaluate the accuracy of the numerical forecast correction model:

[0071]

[0072]

[0073] Among them, Y i is the measured total wind speed value of the ith measured point, is the predicted total wind speed value of the ith measured point, is the average of n true values; the smaller the MAPE, the better the model fitting effect; R 2 It can evaluate the matching degree between the predicted value and the true value, R 2 The closer it is to 1, the better the model effect.

[0074] This embodiment is only a further explanation of the present invention, not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to this embodiment as needed, but as long as it is within the scope of the claims of the present invention, it will be protected by the patent law.

Claims

1. A numerical forecast correction method based on ship-based cruise observation, It is characterized in that The method includes: Step 1: Select a ship-based navigation area, use WRF to perform numerical simulation of the offshore wind field in the ship-based navigation area, and output the WRF numerical simulation results at a time interval of t seconds; Step 2: Calculate the total wind speed uv based on the measured wind speed data of the ship base. The sliding average algorithm is used to calculate the uv and forecast influencing factor data for t seconds of sliding average, and the second-by-second measured data is processed into t-second interval data; Step 3, perform temporal and spatial matching of the t-second interval data processed in step 2 with the WRF numerical simulation results output in step 1 to construct a data set; Step 4: construct a random forest model, divide the constructed data set into a training data set and then a test data set, and use the training data set to train the random forest model to obtain a numerical forecast correction model; Step 5: Input the test data set constructed in step 3 into the numerical forecast correction model constructed in step 4 to test the accuracy of the numerical forecast correction model. If the accuracy test meets the requirements, the numerical forecast correction model is used for the numerical forecast correction of ship-based cruise observations. The average percentage error MAPE and the determination coefficient R 2 Evaluate the accuracy of the numerical forecast correction model: Among them, Y i is the measured total wind speed value of the ith measured point, is the predicted total wind speed value of the ith measured point, is the average of n true values; the smaller the MAPE, the better the model fitting effect; R 2 It can evaluate the matching degree between the predicted value and the true value, R 2 The closer it is to 1, the better the model effect.

2. The method for correcting numerical forecasts based on ship-based cruise observation according to claim 1, It is characterized in that In step 2, abnormal data is pre-eliminated when calculating the total wind speed uv.

3. The method for correcting numerical forecasts based on ship-based cruise observation according to claim 2, It is characterized in that When eliminating abnormal data, the integrity check and internal consistency check of the ship-based measured wind speed data are included.

4. The method for correcting numerical forecasts based on ship-based cruise observation according to claim 1, It is characterized in that In step 2, Among them, u is the radial wind, v is the zonal wind, wspd is the measured wind speed, wdir is the measured wind direction, and uv is the total wind speed.

5. The method for correcting numerical forecasts based on ship-based cruise observation according to claim 1, It is characterized in that In step 3, when performing time-space matching, the longitude, latitude and observation time of the ship-based measured point are used as a reference to extract the data of the point closest to the observation point at the same time in the WRF numerical simulation results.

6. A numerical forecast correction method based on ship-based cruise observation according to claim 5, It is characterized in that In step 2, the forecast influencing factor data include 2-meter temperature, 2-meter relative humidity, 10-meter radial wind u, 10-meter zonal wind v, sea level pressure, and temperature, relative humidity, U, V, and altitude at 925 hPa, 850 hPa, 700 hPa, and 500 hPa.

7. The method for correcting numerical forecasts based on ship-based cruise observation according to claim 6, It is characterized in that The dataset is represented as a two-dimensional matrix: Z m×n ={(x 1,1 ,x 1,2 ,…,x 1,n-1 ,y 1 ),(x 2,1 ,x 2,2 ,…,x 2,n-1 ,y 2 ),…(x i,1 , x i,2 ,…,x i,n-1 ,y i ),(x m,1 ,x m,2 ,…,x m,n-1 ,y m )}, where (x i,1 ,x i,2 ,…,x i,n-1 ,y i ) is the i-th data extracted from the WRF simulation results, i ranges from 1 to m, m is the total number of data, n is the number of characteristic variables, y i is the measured total wind speed value corresponding to the i-th data.

8. The method for correcting numerical forecasts based on ship-based cruise observation according to claim 1, It is characterized in that In step 4, the constructed data set is divided into a training data set and a test data set in a ratio of 4:

1.

9. The method for correcting numerical forecasts based on ship-based cruise observation according to claim 8, It is characterized in that In step 4, building a random forest model includes: Step 4.1, randomly select N training data sets from the training data set by using the bootstrap resampling method, and the N training data sets have M feature attributes; Step 4.2, randomly select a training data set to build, and randomly select m feature attributes from M feature attributes, based on the decision tree model built by the selected training data set, select an optimal feature attribute from the m feature attributes as a partition node, and generate a new decision tree model child node; Step 4.3, repeat step 4.2 until the decision tree model is generated, and build several decision tree models according to steps 4.1-4.3; Step 4.4: Input the training data into all decision tree models, calculate the mean of the prediction values ​​of all decision tree models, and obtain the final prediction result of the random forest model.

Images