A pollen forecasting method based on ensemble learning

Abstract

The present invention relates to the field of pollen forecasting technology, and more specifically, to a pollen forecasting method based on ensemble learning. This method utilizes meteorological numerical model products to construct static features that reflect pollen generation and transmission characteristics, including numerical model features, static features, temporal features, and emission source features, with regional representativeness. This ultimately forms a pollen forecast feature engineering technique. Ensemble learning based on multiple machine or deep learners is employed to improve overall pollen forecasting effectiveness.

Description

Technical Field

[0001] The present invention relates to the technical field of pollen forecasting, and in particular to a pollen forecasting method based on ensemble learning. Background Art

[0002] With the rapid development of the global economy, global warming, and rapid urbanization, pollen concentrations in the air are gradually increasing. The prevalence and severity of hay fever are increasing year by year, posing a serious threat to public health. Pollen forecasting can help pollen allergy sufferers take preventative measures and provide effective early warning information for early intervention and treatment. Therefore, related research has received considerable attention in recent years.

[0003] In the prior art, pollen forecasting mainly includes four modes, namely, prediction based on statistical models, prediction based on machine learning, prediction based on plant growth mechanism and prediction based on transport model. The above four prediction methods all have their own defects, mainly focusing on the lack of information on meteorological factors and the inability to accurately predict extreme values. Among them, the prediction method based on statistical models predicts the pollen yield of plants based on meteorological factors, and then predicts the pollen concentration in the air, but such methods easily ignore the influence of local wind fields, and the climatic conditions in various regions are significantly different, and it is necessary to further consider factors such as turbulence to improve the applicability of different regions. The prediction based on machine learning is similar to the prediction based on statistical models, but meteorological factors are often considered in the construction of features. However, static features such as plant growth status and distribution in different geographical regions have not yet been incorporated into it, and there are fewer extreme value models. The prediction effect for small probability events (pollen extreme values) is not good, which weakens its applicability. The prediction based on plant growth mechanism uses Wuhou models such as temperature forcing or cumulative cooling requirement to make better predictions on the start and end time of pollen, but it cannot predict the specific concentration value of the pollen period, and it is difficult to consider the long-distance transport of pollen. Predictions based on transport models take into account the pollen transport process and incorporate less commonly used factors such as turbulent flux and convection velocity. This can eliminate outliers to a certain extent. However, these models only consider transport, not local pollen emissions. Furthermore, transport varies across different terrains, resulting in poor simulation results in complex terrain. Furthermore, increasing the spatial resolution can lead to distortions in results at high altitudes. Therefore, current pollen prediction models still have significant room for improvement. Summary of the Invention

[0004] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide a pollen forecasting method based on ensemble learning to improve the accuracy of pollen prediction.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] A pollen prediction method based on ensemble learning is provided, which comprises the following steps:

[0007] S1. Use numerical meteorological forecast products to sort out the forecast elements and select meteorological factors that can effectively represent the impact on pollen emission and transport;

[0008] S2. Considering regional differences in pollen emissions, daily NDVI data for different years was obtained through remote sensing data. Static data on height, slope, and aspect at different sites were processed to obtain static features to reflect the differences in pollen emissions across different regions. At the same time, to account for the influence of timeliness, time features such as the day of the year, day of the month, weekday, and month were obtained for each day.

[0009] S3. Arrange the daily pollen concentration of each year into a sequence of annual days in different years, perform smoothing and Gaussian filtering on the sequence, and obtain the multi-year daily average pollen concentration sequence as the pollen emission intensity characteristic;

[0010] S4: Use the results obtained from S1 and S2 as the future daily forecast data. When making the forecast for the current forecast period, use the forecast data from the previous period as one of the features, and add the emission intensity features obtained from S3 to form an integrated learning feature.

[0011] S5. Calculate the feature importance scores of the SHAP importance calculation model, calculate the features obtained in S4, and select features with positive effects as the final integrated learning features;

[0012] S6. Using a machine learning or deep neural network learner, the ensemble learning feature data set obtained in S5 is used for training according to different forecast time limits, and the optimal forecast effect of the learner is obtained as a learning result.

[0013] S7, integrating the primary learning results of each forecast time and the integrated learning features of S5 according to the forecast time to form a new pollen forecast feature, and again using machine learning or deep neural network learners to train according to different forecast time to obtain the optimal forecast effect as the secondary learning result;

[0014] S8, the secondary learning results of different forecast time obtained in S7 are linearly weighted integrated by the least square method according to the forecast time, and finally the pollen daily concentration forecast result of the forecast time is obtained;

[0015] S9. Carry out training and optimization according to multiple forecast time periods, and finally obtain a pollen concentration forecast model that can correspond to the numerical forecast product to realize pollen forecast.

[0016] Furthermore, in step S1, the meteorological factors that can effectively represent the impact on pollen emission and transmission include: basic forecast elements in daily units, basic meteorological characteristics at an altitude of 500 meters above the ground that reflect the impact of transmission, temperature and water vapor factors that control plant growth and thus affect pollen emission, and calculated effective accumulated temperature and accumulated precipitation.

[0017] Furthermore, the basic forecast elements based on daily units include: daily maximum temperature, daily minimum temperature, daily temperature range, daily average wind speed (U), daily average wind speed (V), daily average wind direction, daytime average wind direction, nighttime average wind direction, daily average wind speed, daytime average wind speed, nighttime average wind speed, daily average relative humidity, daytime average relative humidity, nighttime average relative humidity, daily precipitation, daytime precipitation, nighttime precipitation, daily average sea level pressure, daytime average sea level pressure, nighttime average sea level pressure, daily average low cloud cover, daytime average low cloud cover, nighttime average low cloud cover, daily average total cloud cover, daytime average total cloud cover, nighttime average total cloud cover, daily average boundary layer height, daytime average boundary layer height, and nighttime average boundary layer height.

[0018] Furthermore, the basic meteorological characteristics at an altitude of 500 meters above the ground that reflect the impact of transmission include: daily average U wind speed, daily average V wind speed, daily average temperature, daily average relative humidity, daily average wind direction, daily average wind speed, daytime average U wind speed, nighttime average U wind speed, daytime average V wind speed, nighttime average V wind speed, daytime average temperature, nighttime average temperature, daytime average wind speed, nighttime average wind speed, daytime average relative humidity and nighttime average relative humidity.

[0019] Furthermore, in step S2, the specific method for obtaining static data from NDVI data is as follows:

[0020] The daily average gridded NDVI sequence was obtained through filtering and smoothing, and then the station sequence was obtained through bilinear interpolation. Based on the elevation data, the gridded sequence of slope and aspect was calculated, and the static data of height, slope and aspect at different stations were obtained through bilinear interpolation.

[0021] Furthermore, in step S6, machine learning includes: KNN, RandomForest, LinearModel, XGBoost, CatBoost, LightGBM, GBDT, ExtraTrees; deep neural network learners include: multilayer perceptron (MLP) and FastAI.

[0022] The beneficial effects of the present invention are:

[0023] Using meteorological numerical model products, we constructed a set of regionally representative static features that reflect pollen generation and transmission characteristics, including numerical model features, static features, temporal features, and emission source features. This ultimately led to the development of pollen forecast feature engineering technology. We also employed ensemble learning based on multiple machine or deep learners to improve overall pollen forecast performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The multi-time pollen concentration prediction results of the CatBoost model in the embodiment;

[0025] Figure 2 The multi-time pollen concentration prediction results of the ExtraTrees model in the embodiment;

[0026] Figure 3 This is the multi-time pollen concentration prediction result of the KNN model in the embodiment;

[0027] Figure 4 The multi-time pollen concentration prediction results of the model MLP in the embodiment;

[0028] Figure 5 R is the pollen prediction result of multi-machine learning in the embodiment 2 ;

[0029] Figure 6 is the RMSE of the pollen prediction results of multi-machine learning in the embodiment. DETAILED DESCRIPTION

[0030] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0031] Example

[0032] Step 1: Use the European Center numerical forecast products to organize the forecast elements. First, select 45 basic forecast elements on a daily basis, including daily maximum temperature, daily minimum temperature, daily temperature range, daily average U wind speed, daily average V wind speed, daily average wind direction, daytime average wind direction, nighttime average wind direction, daily average wind speed, daytime average wind speed, nighttime average wind speed, daily average relative humidity, daytime average relative humidity, nighttime average relative humidity, daily precipitation, daytime precipitation, nighttime precipitation, daily average sea level pressure, daytime average sea level pressure, nighttime average sea level pressure, daily average low cloud cover, daytime average low cloud cover, nighttime average low cloud cover, daily average total cloud cover, daytime average total cloud cover, nighttime average total cloud cover, daily average boundary layer height, daytime average boundary layer height and nighttime average boundary layer There are a total of 29 ground elements at height; at the same time, in order to reflect the impact of transmission, 16 high-altitude data including daily average U wind speed, daily average V wind speed, daily average temperature, daily average relative humidity, daily average wind direction, daily average wind speed, daytime average U wind speed, nighttime average U wind speed, daytime average V wind speed, nighttime average V wind speed, daytime average temperature, nighttime average temperature, daytime average wind speed, nighttime average wind speed, daytime average relative humidity and nighttime average relative humidity at a height of 500 meters above the ground are added as basic meteorological characteristics; in contrast to the temperature and water vapor factors that affect plant growth and thus pollen emission, the effective accumulated temperature and accumulated precipitation for 3 days, 5 days and 7 days are calculated respectively, for a total of 6 cumulative characteristics; these 51 characteristics will be able to effectively represent the meteorological factors that affect pollen emission and transmission, collectively referred to as numerical model characteristics.

[0033] Step 2: Considering regional variations in pollen emissions, daily NDVI data for different years were obtained using remote sensing data. A daily average gridded NDVI sequence was obtained through filtering and smoothing, and then a station sequence was obtained through bilinear interpolation. Based on global 1km-resolution elevation data, a gridded sequence of slope and aspect was calculated, and static data for elevation, slope, and aspect at each station were obtained through bilinear interpolation. A total of four static features (NDVI, elevation, slope, and aspect) were obtained to reflect regional variations in pollen emissions. To account for the effects of time, four temporal features were obtained: the day of the year, the day of the month, the weekday, and the month.

[0034] Step 3: Arrange the daily pollen concentration of each year into a sequence of annual days in different years, perform smoothing and Gaussian filtering on the sequence, and obtain the multi-year daily average pollen concentration sequence as the pollen emission intensity feature.

[0035] In step 4, in numerical forecasting, the weather conditions for the next few days are forecasted, starting from the current day. The results obtained in steps 1 and 2 are used as the daily forecast data. The forecast data for the next day represents the feature with a forecast timeliness of 1, and so on until the forecast timeliness is 7. Simultaneously, when forecasting the current forecast time (for example, forecast timeliness of 2), the above data from the previous time (forecast timeliness of 1) is used as one of the features, supplemented by the daily emission intensity obtained in step 3, to form the features in the ensemble learning. This results in a total of 51+51+4+4+1=111 features.

[0036] Step 5: Calculate the feature importance score of the model through SHAP importance (SHapley Additive exPlanations Importance), calculate the features obtained in step 4, and select the features with positive effects as the final integrated learning features. The data structure is T×7×N i , i = 1, 2, ..., 111 (where T is the different reporting start time; N is the number of features, which is ultimately determined based on the importance of the features).

[0037] In step 6, a variety of machine learning (KNN, RandomForest, LinearModel, XGBoost, CatBoost, LightGBM, GBDT, ExtraTrees, etc.) and deep neural network learners (multilayer perceptron, FastAI) are used to train the integrated learning feature data set obtained in step 5 according to different forecast time limits and obtain the optimal forecast effect of each learner, which is called a one-time learning result. The result format is T×L×M (where L is the forecast time limit 7; M is 10, indicating the number of learners).

[0038] In step 7, the primary learning results of each forecast time and the integrated learning features of step 5 are integrated according to the forecast time (concatenate, expand the columns of the primary learning results according to the corresponding row index) to form a new pollen forecast feature with a structure of T×L×(111+M). Multiple machine learning and deep neural network learners are used again to train according to different forecast time periods to obtain the optimal forecast effect, which is called the secondary learning result, forming a T×L×M tabular data structure.

[0039] In step 8, the secondary learning results of different forecast time periods obtained in step 7 are linearly weighted integrated using the least squares method according to each forecast time period, and finally the pollen daily concentration forecast result of the forecast time is obtained, with a structure of L×1.

[0040] In step 9, training and optimization are carried out according to the seven forecast time periods, and finally seven pollen concentration forecast models that can correspond to numerical forecast products are obtained.

[0041] This example uses ECMWF real-time forecast data to predict short-term pollen concentrations for a period of 1-3 days, which can provide early warnings for pollen allergy patients and medical departments. This example examines the impact of forecast time on the stability of simulation results. Analysis shows that the multi-time forecast differences of most models are very small, and the fitted lines of pollen forecast concentrations within 3 days are basically overlapping, such as CatBoost ( Figure 1 ), ExtraTrees( Figure 2 ), which means that the multi-time prediction has the same problem as the prediction time of 1. However, the multi-time prediction results of some models showed great differences, which are manifested in the early and sudden appearance of the peak ( Figure 3 , the KNN model showed obvious peak increase and peak burst in mid-to-late March and early April), as well as peak decrease ( Figure 4 , the peak value of the MLP model in mid-to-late March is significantly reduced when ltime=3), and the forecast effect is also reduced and improved accordingly. 2 ( Figure 5 ) and RMSE( Figure 6 ), the forecasting performance of some models decreases with increasing lead time, including LightGBM, RandomForest, CatBoost, ExtraTree, KNN learners, and WeightedEnsemble ensembles. Except for MLP, the remaining models performed relatively stably and were all able to effectively predict short-term pollen concentrations. ExtraTreesMSE and XGBoost performed relatively well.

[0042] In summary, this paper uses meteorological numerical model products to construct static features that reflect pollen generation and transmission characteristics, including numerical model features, static features, temporal features, and emission source features. These features are representative of different regions, ultimately forming a pollen forecast feature engineering technology. This technology utilizes ensemble learning based on multiple machine or deep learners to improve overall pollen forecast performance.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the present invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations that come within the meaning and range of equivalents of the claims be embraced therein.

[0044] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A pollen forecasting method based on ensemble learning, characterized in that: The following steps are involved: S1. Use numerical meteorological forecast products to sort out the forecast elements and select meteorological factors that can effectively represent the impact on pollen emission and transport; S2. Considering regional differences in pollen emissions, daily NDVI data for different years was obtained through remote sensing data. Static data on height, slope, and aspect at different sites were processed to obtain static features to reflect the differences in pollen emissions across different regions. At the same time, to account for the influence of timeliness, time features such as the day of the year, day of the month, weekday, and month were obtained for each day. S3. Arrange the daily pollen concentration of each year into a sequence of annual days in different years, perform smoothing and Gaussian filtering on the sequence, and obtain the multi-year daily average pollen concentration sequence as the pollen emission intensity characteristic; S4: Use the results obtained from S1 and S2 as the future daily forecast data. When making the forecast for the current forecast period, use the forecast data from the previous period as one of the features, and add the emission intensity features obtained from S3 to form an integrated learning feature. S5. Calculate the feature importance scores of the SHAP importance calculation model to calculate the ensemble learning features obtained in S4, and select features with positive effects as the final ensemble learning features; S6. Using a machine learning or deep neural network learner to train the ensemble learning feature dataset obtained in S5 according to different forecast time limits and obtain the optimal forecast effect as a learning result; S7, integrating the primary learning results of each forecast time and the integrated learning features of S5 according to the forecast time to form a new pollen forecast feature, and again using machine learning or deep neural network learners to train according to different forecast time to obtain the optimal forecast effect as the secondary learning result; S8, the secondary learning results of different forecast time obtained in S7 are linearly weighted integrated by the least square method according to the forecast time, and finally the pollen daily concentration forecast result of the forecast time is obtained; S9. Carry out training and optimization according to multiple forecast time periods, and finally obtain a pollen concentration forecast model that can correspond to the numerical forecast product to realize pollen forecast.

2. The pollen prediction method based on ensemble learning according to claim 1, characterized in that: In step S1, the meteorological factors that can effectively represent the impact on pollen emission and transmission include: basic forecast elements on a daily basis, basic meteorological characteristics at an altitude of 500 meters above the ground that reflect the impact of transmission, temperature and water vapor factors that control plant growth and thus affect pollen emission, and calculated effective accumulated temperature and accumulated precipitation.

3. The pollen prediction method based on ensemble learning according to claim 2, characterized in that: The basic forecast elements on a daily basis include: daily maximum temperature, daily minimum temperature, daily temperature range, daily average wind speed (U), daily average wind speed (V), daily average wind direction, daytime average wind direction, nighttime average wind direction, daily average wind speed, daytime average wind speed, nighttime average wind speed, daily average relative humidity, daytime average relative humidity, nighttime average relative humidity, daily precipitation, daytime precipitation, nighttime precipitation, daily average sea level pressure, daytime average sea level pressure, nighttime average sea level pressure, daily average low cloud cover, daytime average low cloud cover, nighttime average low cloud cover, daily average total cloud cover, daytime average total cloud cover, nighttime average total cloud cover, daily average boundary layer height, daytime average boundary layer height, and nighttime average boundary layer height.

4. The pollen prediction method based on ensemble learning according to claim 2, characterized in that: The basic meteorological characteristics at an altitude of 500 meters above the ground that reflect the impact of transmission include: daily average wind speed U, daily average wind speed V, daily average temperature, daily average relative humidity, daily average wind direction, daily average wind speed, daytime average U wind speed, nighttime average U wind speed, daytime average V wind speed, nighttime average V wind speed, daytime average temperature, nighttime average temperature, daytime average wind speed, nighttime average wind speed, daytime average relative humidity and nighttime average relative humidity.

5. The pollen prediction method based on ensemble learning according to claim 1, characterized in that: In step S2, the specific method of obtaining static data from NDVI data is as follows: The daily average gridded NDVI sequence was obtained through filtering and smoothing, and then the station sequence was obtained through bilinear interpolation. Based on the elevation data, the gridded sequence of slope and aspect was calculated, and the static data of height, slope and aspect at different stations were obtained through bilinear interpolation.

6. The pollen prediction method based on ensemble learning according to claim 1, characterized in that: In step S6, machine learning includes: KNN, RandomForest, LinearModel, XGBoost, CatBoost, LightGBM, GBDT, ExtraTrees; deep neural network learners include: multilayer perceptron, FastAI.

Images