Prediction Method of Evapotranspiration and Water Level in Agricultural Irrigation Areas

Abstract

The invention discloses a method for predicting evapotranspiration and water level in an agricultural irrigation area, and belongs to the technical field of agricultural prediction. In order to solve the problem of low prediction accuracy in the existing methods for predicting evapotranspiration and water level in agricultural irrigation areas. The present invention first uses the meteorological data based on each meteorological station and the FAO‑56PM equation to respectively calculate the corresponding evapotranspiration ET0 values; then, a local model, a regional model and a generalized model are respectively established to determine the optimal model for predicting the ET0 value, as the prediction model of ET0. The prediction model of ET0 is then used to predict the future ET0 values, and the decomposition results of VMD decomposition of the water level data and the evapotranspiration meteorological data are combined as the input of the water level prediction model, so as to predict the future water level data.

Description

Technical Field

[0001] The invention belongs to the technical field of agricultural prediction, and in particular relates to a method for predicting evapotranspiration in an agricultural irrigation area and a method for predicting water level in an agricultural irrigation area. Background Art

[0002] Smart agriculture combines modern information science and technology with traditional agriculture, aiming to realize the intelligence, informatization and precision of agricultural production. At present, the field of smart agriculture has made good progress. For example, Attila et al. designed a low-cost, low-power, multi-sensor greenhouse IoT monitoring solution in 2022 to provide agricultural automation monitoring for small farmers. Alves proposed a digital twin application scenario for smart irrigation systems in 2023, which collects and processes soil, weather and crop data through the IoT platform, calculates daily irrigation needs, and evaluates different irrigation strategies. Loversum and Prince conducted an experimental analysis of the machine learning algorithms used in smart agriculture in the same year. However, there are still many aspects that need to be improved in the field of smart agriculture.

[0003] Reference crop evapotranspiration (ET0) refers to the evapotranspiration per unit area of ​​a certain reference crop under certain conditions. ET0 is an important indicator in agriculture, hydrology and environmental science, and is of great significance to agricultural water cycle and decision-making. Machine learning and other algorithms have been used in the estimation of ET0. However, the current evapotranspiration estimation methods still have the problem of low accuracy.

[0004] Water level prediction is one of the hot issues in the research of hydrology, water conservancy and agriculture. In recent years, water conservancy scholars in my country have carried out many studies on water level prediction and achieved certain results. For example, Li Yanfei et al. proposed a hybrid model based on singular spectrum analysis (SSA) method, group method for data processing (GMDH) neural network, weighted integral based on accuracy and diversity (WIAD) and kernel extreme learning machine (KELM) algorithm for predicting river water level. In 2022, Wang Hailin proposed a model for multivariate water level prediction using convolutional radial network, which proved the effectiveness of evapotranspiration data in water level prediction.

[0005] At present, water level prediction research is mainly based on methods such as deep learning and neural networks, constructing a single variable water level prediction model for a single water level data, or considering a multivariate water level prediction model that covers other related factors. Meteorological parameters such as temperature and humidity, rainfall, water flow and evapotranspiration play an important role in improving the accuracy of water level prediction. However, the current water level prediction method has a low prediction accuracy and still needs to be improved. Summary of the invention

[0006] The invention aims to solve the problem of low prediction accuracy in existing agricultural irrigation area evapotranspiration prediction and water level prediction methods.

[0007] A method for predicting evapotranspiration in an agricultural irrigation area uses an ET0 prediction model to predict evapotranspiration in an agricultural irrigation area; the ET0 prediction model is obtained by the following steps:

[0008] S101, obtaining daily meteorological parameters of each meteorological station in the monitoring area within a period of time, the daily meteorological parameters including the year Year and the date day, and the corresponding meteorological parameters, the meteorological parameters including the maximum temperature Tmax and the minimum temperature Tmin, the average relative humidity RH, the wind speed Wind, the sunshine hours SH, the solar radiation flux RAD, and the light intensity LI;

[0009] S102. Feature importance ranking and data division:

[0010] The meteorological data of each meteorological station are recorded as station data sets; at the same time, based on the meteorological data of each meteorological station, the corresponding evapotranspiration ET0 values ​​are calculated using the FAO-56PM equation;

[0011] The meteorological datasets of all stations were combined to form a generalized meteorological dataset;

[0012] The CatBoost algorithm is used to rank the meteorological parameters of each meteorological station. The top three meteorological parameters are selected and the meteorological stations with the same ranking are classified into one region. The obtained regions are recorded as region j, and the data sets of the same region are merged into a regional data set with feature correlation.

[0013] At the same time, the top 4 meteorological parameters in the meteorological parameter importance ranking are taken as input features of the machine learning model; the top 4 meteorological parameters are recorded as characteristic meteorological parameters;

[0014] S103, training local models, regional models and generalized models, and determining the optimal model:

[0015] Select multiple machine learning models Mi as prediction models, where i is the serial number of the machine learning model;

[0016] Based on region j, the station data set corresponding to each station in region j is extracted. Taking each station as the object, characteristic meteorological parameters are extracted from the meteorological parameters as the input features of the prediction model, and the output is the predicted ET0. The ET0 value calculated by the FAO-56PM equation is used as the label to train the prediction model corresponding to each station. The prediction model corresponding to each station is recorded as a local model; at the same time, the test set evaluation indicators MAE, RMSE, R 2 ; For region j, the evaluation index of the local model corresponding to each station in region j is used to determine the optimal machine learning model in region j, which is denoted as MIj and used as the regional model;

[0017] At the same time, the optimal machine learning model in the full range of all regions is determined based on the evaluation indicators of the local models corresponding to each station in all regions, which is recorded as MIJ and used as a generalized model;

[0018] For region j, based on the corresponding regional data set, characteristic meteorological parameters are extracted from the meteorological parameters as the input features of the prediction model, and the regional model MIj is trained. At the same time, the test set evaluation indicators MAE, RMSE, R 2 ;

[0019] At the same time, based on the generalized meteorological data set and the corresponding regional data set, characteristic meteorological parameters are extracted from the meteorological parameters as the input features of the prediction model to train the generalized model MIJ; at the same time, the test set evaluation indicators MAE, RMSE, R 2 ;

[0020] Then, the evaluation index of the regional model is compared with the evaluation index of the local model corresponding to each site, and the regional model or the local model is selected as the proposed prediction model; the evaluation index of the proposed prediction model is compared with the evaluation index of the generalized model as the final prediction model to obtain the ET0 prediction model;

[0021] In the process of using the ET0 prediction model to predict the evapotranspiration of agricultural irrigation areas, the characteristic meteorological parameters obtained based on the meteorological parameters are first input into the ET0 prediction model for ET0 prediction.

[0022] Furthermore, the characteristic meteorological parameters include Tmax, RH, Wind and LI.

[0023] Furthermore, the machine learning models in S103 include XGboost, Random Forest RF and Catboost.

[0024] A method for predicting evapotranspiration in an agricultural irrigation area uses an ET0 prediction model to predict evapotranspiration in an agricultural irrigation area; the ET0 prediction model is obtained by the following steps:

[0025] A101. Obtain daily meteorological parameters of each meteorological station in the monitoring area within a period of time, where the daily meteorological parameters include the year Year and the date day, and corresponding meteorological parameters, where the meteorological parameters include the maximum temperature Tmax and the minimum temperature Tmin, the average relative humidity RH, the wind speed Wind, the sunshine hours SH, the solar radiation flux RAD, and the light intensity LI;

[0026] A102. Feature importance ranking and data division:

[0027] The meteorological data of each meteorological station are recorded as station data sets; at the same time, based on the meteorological data of each meteorological station, the corresponding evapotranspiration ET0 values ​​are calculated using the FAO-56PM equation;

[0028] The CatBoost algorithm is used to rank the meteorological parameters of each meteorological station. The top three meteorological parameters are selected and the meteorological stations with the same ranking are classified into one region. The obtained regions are recorded as region j, and the data sets of the same region are merged into a regional data set with feature correlation.

[0029] At the same time, the top 4 meteorological parameters in the meteorological parameter importance ranking are taken as input features of the machine learning model; the top 4 meteorological parameters are recorded as characteristic meteorological parameters;

[0030] A103. Based on region j, characteristic meteorological parameters are extracted from the corresponding regional data set as input features of the prediction model, and the output is the predicted ET0. The ET0 value calculated by the FAO-56PM equation is used as the label to train the prediction model corresponding to region j. Multiple machine learning models are trained for each region as prediction models, and the test set evaluation indicators of the model, such as MAE, RMSE, and R are recorded. 2 , and then select the optimal machine learning model as the ET0 prediction model corresponding to region j based on the evaluation index;

[0031] In the process of using the ET0 prediction model to predict the evapotranspiration of agricultural irrigation areas, the characteristic meteorological parameters are first obtained based on the meteorological parameters, and the ET0 prediction model of the area where the characteristic meteorological parameters correspond to the station is input to perform ET0 prediction.

[0032] Furthermore, the characteristic meteorological parameters include Tmax, RH, Wind and LI.

[0033] Furthermore, the machine learning models in S103 include XGboost, Random Forest RF and Catboost.

[0034] A method for predicting water level in an agricultural irrigation area comprises the following steps:

[0035] S201, obtaining historical water level data and evapotranspiration meteorological data, wherein the evapotranspiration meteorological data includes maximum temperature Tmax, minimum temperature Tmin, relative humidity RH and evapotranspiration ET0;

[0036] S202, using fast Fourier transform to analyze the water level data, obtain the frequency result of the main peak of the water level data, and then determine the time series sliding window length N of the time series model based on the frequency result of the main peak of the water level data;

[0037] S203, using VMD variational mode decomposition algorithm to perform VMD decomposition on the water level data to obtain 8 decomposition signals IMF1-IMF8;

[0038] Based on IMF1~IMF8 and evapotranspiration meteorological data, the Pearson correlation coefficient method is used to determine the three IMFs in IMF1~IMF8 with high correlation coefficients with water level data, and the three IMFs, evapotranspiration meteorological data and water level data, a total of 8 features, constitute multivariate water level prediction data;

[0039] S204, training water level prediction model:

[0040] Based on 8 historical features, multivariate water level prediction data and a time series sliding window length N are formed, a feature matrix of (N, 8) is constructed as the input of a water level prediction model, and the predicted value of the future water level is used as the output of the water level prediction model to train the water level prediction model and obtain a trained water level prediction model; the neural network of the water level prediction model includes a one-dimensional convolutional neural network CNN layer and a GRU network set after the one-dimensional convolutional neural network CNN layer;

[0041] S204, predicting the future water level using the trained water level prediction model; when predicting the future water level, the future evapotranspiration data in the input data is the evapotranspiration data predicted by using the agricultural irrigation area evapotranspiration prediction method.

[0042] Further, the time series sliding window length N of the time series model is determined based on the period of the main peak 3 analyzed using fast Fourier transform.

[0043] Furthermore, the length N of the time series sliding window is 35 days.

[0044] Further, three IMFs among IMF1 to IMF8 having high correlation coefficients with the water level data are IMF1, IMF6, and IMF7.

[0045] Beneficial effects:

[0046] 1. Compared with the reference ET0 value calculated by the Penman equation, the agricultural irrigation area evapotranspiration prediction scheme using the regional model in the present invention has stronger prediction stability and higher accuracy, and only four meteorological parameters are needed to achieve accurate estimation of evapotranspiration ET0, reducing the cost of evapotranspiration calculation.

[0047] 2. The multivariate VMD-CNN-GRU model integrating evapotranspiration meteorological data proposed in the present invention has high prediction accuracy in multi-step water level prediction, and the overall error offset is more stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is the distribution map of meteorological stations in Heilongjiang Province.

[0049] Figure 2 Flowchart for designing an evapotranspiration estimation model.

[0050] Figure 3 This is a feature importance ranking graph.

[0051] Figure 4 A distribution map of each meteorological station area is provided.

[0052] Figure 5 This is a comparison chart of the ET0 calculated values ​​of the region 1 model.

[0053] Figure 6 This is a comparison chart of the ET0 calculated values ​​of the region 2 model.

[0054] Figure 7 This is a comparison chart of ET0 calculated values ​​of the three regional models.

[0055] Figure 8 This is the VMD mode decomposition diagram (Harbin dataset).

[0056] Fig. 9 This is the VMD mode decomposition diagram (Qitaihe data set).

[0057] Fig.10 This is the structure diagram of the CNN-GRU model.

[0058] Fig.11 Construct an overall flow chart for the model.

[0059] Fig.12 This is a comparison chart between multivariate and univariate models (Harbin).

[0060] Fig.13 This is a comparison chart between multivariate and univariate models (Qitaihe).

[0061] Fig.14 Predict the RMSE change trend of step length for each model (Harbin dataset).

[0062] Fig.15 Prediction trend of RMSE step length for each model (Qitaihe dataset).

[0063] Fig.16 The following is a performance comparison chart of the three local models in each area. DETAILED DESCRIPTION

[0064] The present invention aims to analyze the contribution of each meteorological parameter to the evapotranspiration estimation of a province, and help select the sensor type of the Internet of Things monitoring terminal, design an evapotranspiration estimation model, and then combine the multivariate water level prediction model of multiple variables such as water level data internal information and meteorological parameters to analyze the contribution of each meteorological parameter to the evapotranspiration estimation of Heilongjiang Province, and design the sensor type of the Internet of Things monitoring terminal to further improve the water level prediction accuracy. The present invention is described in detail below in conjunction with specific embodiments. Specific implementation method one:

[0066] This implementation is a method for predicting evapotranspiration in agricultural irrigation areas. By analyzing the characteristic relationship between evapotranspiration and meteorological parameters, high-importance meteorological parameters are screened out, providing a basis for the water level prediction system in agricultural irrigation areas. This implementation designs an evapotranspiration estimation model based on the regional division feature of machine learning. Only a small amount of meteorological data is needed to complete the estimation of evapotranspiration, which greatly improves the time efficiency. At the same time, before building the model, CatBoost is used to sort the importance of meteorological data to screen effective features and further improve the model efficiency. Moreover, a regional data set is constructed by dividing it into three regions, and a regional evapotranspiration estimation model is established to address the shortcomings of existing generalized models and regional model methods.

[0067] The method for predicting evapotranspiration in an agricultural irrigation area described in this embodiment comprises the following steps:

[0068] S101, obtain data and perform preprocessing:

[0069] The meteorological data of the present invention include daily meteorological parameters of 25 meteorological stations in Heilongjiang Province from 1960 to 2015, including the maximum temperature (Tmax) and the minimum temperature (Tmin), the average relative humidity (RH), the wind speed at a height of 2 meters (Wind), the sunshine hours (SH), the solar radiation flux (RAD), the light intensity (LI), the year and the day (Year, day);

[0070] Map of Heilongjiang Province and the distribution of 25 weather stations Figure 1 These data are used to calculate the standard reference evapotranspiration value (ET0), which is the prediction target of the subsequent machine learning algorithm. The calculation formula uses the Penman-Monteith equation (FAO-56PM) certified by the Food and Agriculture Organization of the United Nations (FAO). Some of the data are shown in Tables 1 and 2.

[0071] Table 1 Statistics of 25 meteorological stations

[0072]

[0073] Table 2 Daily meteorological data (taking Harbin dataset as an example)

[0074]

[0075] In terms of evapotranspiration estimation, the present invention has conducted a lot of research on machine learning methods, mainly generalized models and local models. The generalized model combines the data of multiple meteorological stations into a large data set for analysis and training, and selects the model with the best effect as the model for the entire region; the local model is trained independently for the data set of each meteorological station, and the most suitable model for the meteorological station is selected by comparing the performance of each model.

[0076] These two methods have their own advantages and disadvantages. The advantage of the generalized model is that the data set is larger, which can improve the adaptability and robustness of the model to unknown data and enhance the stability of the model. However, the training time of the generalized model is long, the feature selection is difficult, and the differences between different regions cannot be considered, which may lead to deviations in evapotranspiration estimation. The advantages of the local model are that the training data volume is small, the training time is short, and it is targeted to the local area. However, the local model is prone to overfitting, poor generalization performance, and inaccurate feature selection, which affects the model performance.

[0077] In order to solve the above problems, the present invention proposes a method for dividing meteorological stations according to characteristic regions. That is, the data sets of multiple meteorological stations are divided into different combinations, the meteorological data (time, wind speed, humidity, maximum temperature, minimum temperature, light intensity and sunshine duration) of each meteorological station are sorted by feature importance, and the top 3 meteorological parameters are taken and recorded. Meteorological stations with the same ranking are classified into one type of area, and the data sets are merged to form a regional data set with feature correlation to train the model, that is, the regional model. At the same time, the geographical distribution of each meteorological station is referred to to judge the rationality of the division result.

[0078] The overall process of the ETO algorithm proposed in the present invention is as follows: Figure 2 As shown. The data preprocessing mainly uses interpolation to fill in abnormal values ​​and missing values, and normalizes the data. The normalization formula is as follows:

[0079]

[0080] The preprocessed data should be divided into training data and test data. The training data is used to train the corresponding machine learning models, and the test data is used to test the generalization performance of the machine learning models.

[0081] S102. Feature importance ranking and data division:

[0082] When using the FAO-56PM equation to calculate the ET0 value, more meteorological parameters are needed. In addition, for calculating evapotranspiration using the FAO-56PM equation, each parameter is equally important and indispensable in the calculation process, which further leads to the problem of too many parameters.

[0083] The present invention uses the CatBoost algorithm to sort the meteorological data of 25 meteorological stations by feature importance. The feature importance result values ​​(retain two decimal places) are shown in Table 3, and the feature sorting stack is shown in Figure 3 shown.

[0084] Table 3 Catboost feature importance ranking results

[0085]

[0086]

[0087] According to the results of feature sorting, it can be found that at all meteorological stations, the maximum temperature (Tmax) has a very high feature importance, which shows that the meteorological parameter Tmax has a very important influence on the calculation of evapotranspiration in the entire Heilongjiang region. In the literature description of most evapotranspiration-related studies, Tmax is also the meteorological parameter with the greatest influence on evapotranspiration. At the same time, at most stations, Tmax, RH, Wind and LI have high feature importance, so the present invention uses these four meteorological parameters as input features of the evapotranspiration estimation model.

[0088] Then, for each station, the top three features in terms of importance are selected. The bold fonts in Table 2 are the top three features with the highest importance. Stations with the same top three features are classified into the same area. In this way, we divide the 25 meteorological stations into three types of areas. The division results are shown in Tables 4 and Figure 4 shown.

[0089] Table 4 Regional distribution of meteorological stations

[0090]

[0091]

[0092] According to the results of feature division, the meteorological data of the meteorological stations in the three regions were merged to construct three regional meteorological data sets, which were used to train their corresponding regional models. At the same time, the data of all 25 meteorological stations were merged to form a generalized meteorological data set to train the generalized model.

[0093] When establishing a model, the present invention selects the first four meteorological parameters with higher feature importance, namely, maximum temperature, relative humidity, wind speed, and light intensity, as input features for regional model and generalized model training according to the sorted feature importance results. Other parameters are discarded because of their lower importance and poor contribution to the model, which can also improve the speed of model training.

[0094] At the same time, according to the results of the above experimental data, it can be found that the meteorological stations divided into regions one and three are geographically adjacent, which is consistent with the fact that in places with close geographical locations, the climate environment is similar, and their meteorological data have the same machine learning feature sorting results; the meteorological stations divided in region two are discretely distributed and have no geographical adjacent relationship. They are divided into one type of region here because the data sets of these meteorological stations have similar performance results in the machine learning sorting algorithm. In the subsequent machine learning model training, the overall performance of the model in region two is also better than the generalized model and the local model, which also shows that machine learning methods can find deeper correlation relationships.

[0095] S103. Build and train local models, regional models and generalized models and conduct comparative experiments:

[0096] The meteorological data of the three regions containing meteorological stations were trained using three algorithms, including XGboost, random forest RF and Catboost, and the meteorological data of each meteorological station in each region were used to establish a machine learning model. Specifically, we used the 12 meteorological stations included in region one to establish 12 local models. For the convenience of subsequent explanations, these 12 local models are referred to as region one local models. The establishment process of region two local models and region three local models is the same as that of region one local models. These models are collectively referred to as regional local models. The input of the regional local model is the meteorological data of the meteorological station, and the output is the estimated evapotranspiration ET0. The evaluation indicators used are MAE, RMSE, and R 2 Then calculate the MAE, RMSE, and R of the local model in each region 2 The average value of is compared and the algorithm with the best overall performance in each region is selected as the regional model algorithm.

[0097] Table 5 Regional-local model training results

[0098]

[0099]

[0100] Table 6. Local model training results for region 2

[0101]

[0102] Table 7. Training results of three local models in the region

[0103]

[0104] The performance comparison of regional local models is as follows Fig.16 As shown, (a), (b), and (c) are the R of region 1, region 2, and region 3, respectively. 2Curve (d), (e), (f) are RMSE curves of region 1, region 2, and region 3, respectively; (g), (h), (i) are MAE curves of region 1, region 2, and region 3, respectively. The analysis of the experimental results shows that in region 1, the Catboost algorithm model performs best. In region 2, the Xgboost algorithm model performs best. In region 3, the best performing model is RF. It can also be found that the Catboost algorithm performs better on a large number of training models, and the performance of the Catboost algorithm is relatively stable on most meteorological stations (as shown by Fig.15 Therefore, the present invention adopts Catboost algorithm to train the generalized model.

[0105] The best algorithm for each region has been determined. Next, the regional site data fusion training is carried out. The steps are as follows:

[0106] (1) All site data sets in region 1 are merged, and the model for region 1 is trained using the Catboost algorithm;

[0107] (2) All site data in region 2 are merged, and the region 2 model is trained using the XGboost algorithm;

[0108] (3) All site data in region 3 are merged, and the RF algorithm is used to train the region 3 model;

[0109] (4) The data of all 25 sites were merged to form a generalized dataset, and the Catboost algorithm was used to train the generalized model.

[0110] The training results and indicators of the four models are shown in Table 8 and Figure 5 , Figure 6 , Figure 7 shown.

[0111] Table 8 Training results of each model

[0112]

[0113] From the comparative analysis between the regional model and the generalized model, it can be found that the regional model has a better effect on evapotranspiration estimation than the generalized model. By comparing the mean values ​​of the evaluation indicators of the local model of region 1, the local model of region 2, and the local model of region 3 in Table 8 and Table 5, Table 6, and Table 7, it can be found that the performance of the three regional models is also better than that of the local models corresponding to each station in the region, which also verifies the effectiveness of the machine learning model method for regional division. At the same time, it can also be found that in terms of the performance of each data set in region 2, the R of the model for region 2 established by the present invention is 2 The coefficient is 0.9891, while the average R 2The coefficient is 0.9645, which shows that the overall model of region two is better than the local model of region two, and also verifies the feasibility of dividing regions according to the ranking of machine learning features.

[0114] Measured data verification:

[0115] The above has completed the regional division of 25 meteorological stations in Heilongjiang Province, and established the corresponding optimal machine learning model, which was verified using measured data to compare the effect of the verification model. The measured data is the data from July 1, 2022 to October 1, 2023 in Harbin collected by the hardware monitoring terminal and uploaded to the Alibaba Cloud IoT platform. At the same time, in order to verify the comparison, it is necessary to calculate the reference ET0 value. The data for calculating the reference evapotranspiration comes from the wheat malt-agricultural big data system software. The meteorological data of Harbin during the same period was obtained as the parameter for calculating the reference ET0 of the FAO-56PM equation. The regional model selected is the second model of the Harbin region.

[0116] The above content shows that the performance of the regional model is better than that of the regional local model, which further verifies the effectiveness of dividing the region according to the machine learning feature ranking. In order to further verify the estimation accuracy of the evapotranspiration estimation model proposed in this invention, on the basis of the above experiments, we use measured data for further verification. The measured data is the data from July 1, 2022 to October 1, 2023 in Harbin collected by the hardware monitoring terminal and uploaded to the Alibaba Cloud IoT platform.

[0117] After experimental comparison, the ET0 value calculated by the regional model is closer to the reference ET0 value, the ET0 value fluctuates greatly, the ET0 value calculated by the generalized model has a large error, the ET0 value calculated by the regional model is closer to the standard reference value, and the model performance is more stable. It verifies that the regional model is more accurate and targeted in predicting evapotranspiration in the region.

[0118] The above process is to input the station data in area 2 into the model corresponding to area 2 where the station is located for prediction. For stations in other areas, the meteorological data of the station should be input into the regional model corresponding to the area where the station is located for prediction.

[0119] The above experiment shows that compared with the reference ET0 value calculated by the Penman equation, the regional model constructed by the present invention has stronger prediction stability and higher accuracy, and only four meteorological parameters are needed to accurately estimate the evapotranspiration ET0, reducing the cost of evapotranspiration calculation. In terms of model deployment, it is only necessary to match the nearest meteorological station based on the GPS longitude and latitude information collected by the hardware monitoring terminal to determine the region to which it belongs, and then call the corresponding regional model to estimate the ET0 value. Specific implementation method 2:

[0121] This embodiment is a method for predicting water levels in agricultural irrigation areas. Based on the evapotranspiration estimation method of the first embodiment, the method for predicting water levels in agricultural irrigation areas is implemented. The method for predicting water levels in agricultural irrigation areas described in this embodiment has the following characteristics:

[0122] 1. The traditional water level prediction method is heavily dependent on manual experience. The present invention uses a deep time series neural network to solve this problem. Specifically, the present invention uses LSTM, BiLSTM and GRU three time series models as the backbone network of the univariate water level prediction model, and obtains the best hyperparameters through SSA hyperparameter optimization. The subsequent multivariate water level prediction model also uses the same hyperparameters. Then, the best time series model is selected through experimental comparison as the backbone network of the multivariate water level prediction model.

[0123] 2. The time series sliding window is an important hyperparameter of the deep time series neural network. Existing studies only use manual experience and hyperparameter search, which not only makes the selection of the time series sliding window unscientific and uninterpretable, but also leads to serious resource consumption and poor interpretability. Therefore, the present invention proposes a method for selecting the optimal time series sliding window size using fast Fourier transform.

[0124] 3. In the preprocessing of raw water level data, existing studies only focus on the impact of missing values, while ignoring the importance of outliers. Therefore, inspired by the outlier processing methods in other fields, this paper proposes to use an outlier detection algorithm to detect outlier data in water level data, and then use interpolation method for subsequent processing.

[0125] 4. Existing research is weak in mining the potential information of single water level data, and the water level data has nonlinear and unstable characteristics, which will affect the final prediction performance of the model. In view of this, the VMD variational mode decomposition algorithm of the present invention further mines the information contained in the single water level data and eliminates the nonlinear and unstable characteristics of the water level data, further improving the prediction performance of the model. After processing, 8 groups of mode decomposition features such as IMF1 to IMF8 are obtained, which are components of the multivariate water level prediction data of the present invention.

[0126] 5. The VMD variational modal algorithm requires two parameters to be manually determined. Existing studies use manual determination of these two parameters, which may cause errors due to human experience. Therefore, the present invention proposes to use an artificial gorilla troops optimizer (AGTA) to optimize the parameters of the VMD algorithm.

[0127] 6. Existing studies have shown that the factors that affect water levels are diverse. Inspired by this, the present invention further combines evapotranspiration, meteorological data, modal decomposition features after decomposition using the VMD algorithm, and water level data to form preliminary multivariate water level prediction data; other researchers have shown that not all input features have a promoting effect on the model. Inspired by this, the present invention uses the random forest importance ranking algorithm and the Pearson correlation coefficient method to further analyze the preliminary multivariate water level prediction data, and finally constructs multivariate water level data after screening and elimination.

[0128] 7. Existing studies directly input multiple variables into the model without considering the problem of increased model complexity caused by the increase in variables. Therefore, the present invention introduces a one-dimensional convolutional layer before the time series model. By introducing the one-dimensional convolutional layer, it is possible to locally perceive the input data and capture the local feature patterns in the time series, thereby achieving a more efficient prediction model and finally constructing a VMD-CNN-GRU multivariate water level prediction model.

[0129] More specifically, a method for predicting water level in an agricultural irrigation area described in this embodiment includes the following steps:

[0130] S201, combining internal information of water level data, meteorological data and evapotranspiration to determine multivariate water level data for a water level prediction model:

[0131] The water level data comes from the Hongxing Reservoir in Harbin and the Taoshan Reservoir in Qitaihe, covering the water level data from 2010 to 2015. The data is collected from the National Water Information Website of the Information Center of the Ministry of Water Resources and the Harbin Public Dataset Open Platform. The data includes the daily changes in reservoir water levels.

[0132] The meteorological data include the daily meteorological parameters of Harbin and Qitaihe meteorological stations from 2010 to 2015. The daily meteorological parameters include the maximum temperature (Tmax) and the minimum temperature (Tmin), average relative humidity (RH), wind speed at a height of 2 meters (Wind), sunshine hours (SH), solar radiation flux (RAD), light intensity (LI), year and day (Year, day).

[0133] Based on the water level prediction in the prior art, the present invention further combines the internal information of water level data, evapotranspiration, meteorological data and water level data, and further combines the evapotranspiration, meteorological data and water level data to form multivariate water level prediction data;

[0134] The evapotranspiration of Harbin and Qitaihe is estimated according to the meteorological data and the evapotranspiration estimation method described in Specific Implementation Method 1. Here, only four meteorological characteristics, namely maximum temperature, relative humidity, wind speed, and light intensity, are used to calculate the evapotranspiration. The evapotranspiration is then combined with the meteorological data to form evapotranspiration meteorological data, and preprocessing such as data outliers, missing values, and normalization is performed; the evapotranspiration meteorological data is then sorted using random forest importance to screen out the most effective variables, and the top five characteristics in importance ranking are Tmax, Tmin, ET0, RH, and LI; the Pearson correlation coefficient is then used to analyze the correlation between these five characteristics and the water level data, and it is concluded that the water level data and evapotranspiration data have a certain degree of correlation, among which the maximum temperature, minimum temperature, relative humidity, and evapotranspiration value are highly correlated with the water level, so these four meteorological parameters are selected in combination with the water level as new features of the water level prediction model;

[0135] S202. Since the time series sliding window is an important hyperparameter of the deep time series neural network, and the existing research only uses manual experience and hyperparameter search, this will not only make the selection of the time series sliding window unscientific and unexplainable, but also lead to serious resource consumption and poor interpretability. Compared with the hyperparameter search, the fast Fourier transform requires much fewer resources. At the same time, the fast Fourier transform analyzes the periodicity of the water level data and takes the actual situation into consideration, which is more scientific than manual experience. In summary, the present invention uses the fast Fourier transform to analyze the water level data and determine the optimal time series sliding window more scientifically and efficiently. Specifically, we first obtain the frequency results of the first 8 main peaks of the water level data through the fast Fourier transform, as shown in Table 9.

[0136] Through the main frequency results in Table 9, it can be found that the period of main peak 1 is greater than 100 days, which reflects the overall transformation period of water level data, that is, the trend of new seasonal changes in water level. Main peaks 2 to 8 reflect the fluctuation period of data in a small range, among which the amplitude of main peak 3 is the largest, so the present invention selects main peak 3 as the focus of subsequent analysis. In the Harbin data set, the period of main peak 3 is 34.7 days, and in the Qitaihe data set it is 33.1 days. The present invention finally selects 35 days as the time series sliding window of the time series model. The specific reasons are as follows: Considering the principle of periodic consistency, selecting a time series sliding window size that is closer to these two periods can better capture these periodic features. By selecting 35 days as the time series window size, the time series information contained in the two data sets is retained to a certain extent. On the other hand, a larger time series window will provide more stable input feature information. The method of the present invention comprehensively considers the actual situation, is more scientific, and has less resource consumption. Therefore, the present invention adopts a time series sliding window size of 35 days, that is, inputs the data of the first 35 days to predict the future water level.

[0137] Table 9 Cycle search results

[0138]

[0139]

[0140] S203. Since existing research is weak in mining the potential information of single water level data, and the water level data has nonlinear and unsteady characteristics, which will affect the final prediction performance of the model, in view of this, the present invention adopts the VMD variational mode decomposition algorithm to further mine the information contained in the single water level data and eliminate the nonlinearity and unsteadiness of the water level data, so as to further improve the prediction performance of the model. When performing VMD decomposition, it is necessary to manually set the penalty factor and the modal decomposition number. The traditional method relies on manual experience to determine, but this method is not scientific. The present invention uses the Artificial Gorilla Troops Optimizer (AGTA) to optimize the VMD parameters. The optimized parameters are the decomposition mode number of 8 and the penalty factor of 695. VMD decomposition is performed according to the optimized parameters, and the decomposed signal waveform obtained is as follows: Figure 8 and Fig. 9 After VMD decomposition, 8 modal decomposition features (IMF1 to IMF8) are obtained as auxiliary features for water level prediction.

[0141] Next, the VMD modal decomposition algorithm was used to obtain eight modal decomposition features, including IMF1 to IMF8, on the water level data. The features were merged with the four evapotranspiration meteorological data (maximum temperature, minimum temperature, relative humidity and evapotranspiration) selected above, and further analyzed using the Pearson correlation coefficient method. It was found that the correlation coefficient between the evapotranspiration meteorological data and the water level data was higher than that of the IMF modal decomposition features as a whole. This also verified the feasibility and effectiveness of combining evapotranspiration data with digital data to predict water levels. At the same time, among the eight VMD modal decomposition features, IMF1, IMF6 and IMF7 had higher correlations. Therefore, IMF1, IMF6 and IMF7 were selected together with the evapotranspiration meteorological data and water level data to form the multivariate water level prediction data.

[0142] S204. Build a multivariate water level prediction model based on VMD-CNN-GRU:

[0143] The present invention studies water level prediction from two aspects: one is to directly input water level data as features into a time series model, and the model output is the future water level prediction value, that is, a univariate water level prediction model; the other is to use the VMD algorithm to conduct an in-depth analysis of the water level data, and further form multivariate water level prediction data by combining the three modal decomposition features finally obtained with evapotranspiration meteorological data to form a multivariate water level prediction model. The model input is a (35,8) matrix, where the first dimension 35 indicates that the time sliding window size is 35 days, and the second dimension 8 indicates the features of 8 multivariate water level prediction data, and the model output is also the future water level prediction value. The (35,8) matrix is ​​actually a 35*8 matrix, and each row represents 8 features corresponding to one day.

[0144] First, the univariate water level prediction model is analyzed, and three time series prediction models, LSTM, BiLSTM and GRU, are used. SSA hyperparameter optimization is used to finally establish a univariate water level prediction model. The results show that the performance of the GRU model is better than the other two time series models in univariate prediction, so the GRU model is also used in the establishment of the multivariate water level prediction model.

[0145] In order to cope with the increased complexity caused by the increase in the number of features, considering that the multivariate water level prediction data has 8 features, compared with using only water level data, the multivariate water level prediction data has a high dimension and a complex spatiotemporal structure. Therefore, the present invention introduces a one-dimensional convolutional neural network (CNN) layer before the multivariate GRU model. The advantages of introducing a one-dimensional convolutional layer are twofold: on the one hand, by designing the number of channels (number of filters) of the convolutional layer, the feature dimension is increased without increasing the receptive field, and the high-dimensional characteristics of the feature meteorological data are better captured, while making the network structure deeper; on the other hand, the one-dimensional convolutional layer better captures the periodic laws and trends of evapotranspiration meteorological data and water level time series data. Fig.10 The structure of the multivariate water level prediction model designed by the present invention is shown. Fig.11 The overall process of model construction is demonstrated.

[0146] Based on the univariate GRU model, a multivariate GRU water level prediction model was constructed, and the same SSA optimization algorithm was used for hyperparameter optimization. Finally, a VMD-CNN-GRU multivariate water level prediction model integrating evapotranspiration meteorological data was established and compared with the univariate GRU model. The experimental results are shown in Table 10. Fig.12 and Fig.13 A comparison of the performance of univariate and multivariate water level predictions is shown.

[0147] Table 10 Evaluation table of multivariate GRU model and univariate GRU model

[0148]

[0149] From Table 10 and Fig.12 , Fig.13 The experimental results show that the multivariate water level prediction model VMD-CNN-GRU, which combines evapotranspiration-related meteorological data and VMD decomposition modal data, performs better than the univariate water level prediction model SSA-GRU in single-step water level prediction. Since the GRU model has been proven to be better than the LSTM and BiLSTM models in univariate prediction, the present invention only uses the GRU model when constructing a multivariate model. It is verified here that the multivariate water level prediction model constructed by the GRU model is better than the univariate prediction of the GRU. From the perspective of machine learning and deep learning, the multivariate model introduces more data than the univariate model, provides more reference basis for model prediction, increases the number of effective input features, and improves the prediction effect of the model accordingly. However, it can be found from Table 10 that the multivariate model is better than the univariate model in R 2 The coefficient increased from 0.982 to 0.989, which is not a large increase. Other indicators are also similar. Therefore, in order to further compare the performance difference between the multivariate model and the univariate model, the present invention sets up a comparative experiment of multi-step prediction.

[0150] When predicting the future water level, the future evapotranspiration data in the input data is the evapotranspiration data predicted by using the agricultural irrigation area evapotranspiration prediction method described in the first specific implementation mode.

[0151] The above experiments are all single-step water level predictions, that is, using historical water level data to predict the water level forecast value for one step or one day in the future. In contrast, multi-step water level prediction refers to predicting the water level change trend within multiple time steps in the future based on historical water level data, which has better guiding significance. However, a prediction step that is too long will also make the prediction results of the model inaccurate. Based on the single variable water level prediction model in this section, a multi-step water level prediction experiment was carried out. The prediction results of the three single variable water level prediction models and the multivariable water level prediction model proposed in this invention were compared at 10 steps, and the root mean square error (RMSE) was used as a reference indicator for evaluation. The experimental results are shown in Tables 11-12 and Figure 14-15 .

[0152] Table 11 Comparison of RMSE of different step length predictions (Harbin dataset)

[0153]

[0154] Table 12 Comparison of RMSE of different step length predictions (Qitaihe dataset)

[0155]

[0156]

[0157] According to Table 11, Table 12 and Fig.14 , Fig.15 From the experimental results, it can be found that the multivariate VMD-CNN-GRU model that integrates evapotranspiration meteorological data proposed in the present invention has better overall prediction effects in multi-step prediction of water level than the three univariate prediction model methods, and the overall error offset is more stable. The experiment also reflects that in single-step prediction, there is not much difference between the univariate model and the multivariate model, but in multi-step prediction, the effect of the multivariate model is obviously better than the univariate model. Therefore, it can be concluded that the VMD-CNN-GRU model that integrates meteorological data proposed in the present invention has better prediction accuracy in water level prediction than the univariate model, and performs more steadily in multi-step prediction. At the same time, it can also be found that among the three univariate models, GRU also has the best multi-step prediction performance.

[0158] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting evapotranspiration in agricultural irrigation areas, characterized in that: Use ET0 prediction model to predict evapotranspiration in agricultural irrigation areas; The ET0 prediction model is obtained by the following steps: S101, obtaining daily meteorological parameters of each meteorological station in the monitoring area within a period of time, the daily meteorological parameters including the year Year and the date day, and the corresponding meteorological parameters, the meteorological parameters including the maximum temperature Tmax and the minimum temperature Tmin, the average relative humidity RH, the wind speed Wind, the sunshine hours SH, the solar radiation flux RAD, and the light intensity LI; S102. Feature importance ranking and data division: The meteorological data of each meteorological station are recorded as station data sets; at the same time, based on the meteorological data of each meteorological station, the corresponding evapotranspiration ET0 values ​​are calculated using the FAO-56PM equation; The meteorological datasets of all stations were combined to form a generalized meteorological dataset; The CatBoost algorithm is used to rank the meteorological parameters of each meteorological station. The top three meteorological parameters are selected and the meteorological stations with the same ranking are classified into one region. The obtained regions are recorded as region j, and the data sets of the same region are merged into a regional data set with feature correlation. At the same time, the top 4 meteorological parameters in the meteorological parameter importance ranking are taken as input features of the machine learning model; the top 4 meteorological parameters are recorded as characteristic meteorological parameters; S103, training local models, regional models and generalized models, and determining the optimal model: Select multiple machine learning models Mi as prediction models, where i is the serial number of the machine learning model; Based on region j, the station data set corresponding to each station in region j is extracted. Taking each station as the object, characteristic meteorological parameters are extracted from the meteorological parameters as the input features of the prediction model, and the output is the predicted ET0. The ET0 value calculated by the FAO-56PM equation is used as the label to train the prediction model corresponding to each station. The prediction model corresponding to each station is recorded as a local model; at the same time, the test set evaluation indicators MAE, RMSE, R 2 ; For region j, the evaluation index of the local model corresponding to each station in region j is used to determine the optimal machine learning model in region j, which is denoted as MIj and used as the regional model; At the same time, the optimal machine learning model in the full range of all regions is determined based on the evaluation indicators of the local models corresponding to each station in all regions, which is recorded as MIJ and used as a generalized model; For region j, based on the corresponding regional data set, characteristic meteorological parameters are extracted from the meteorological parameters as the input features of the prediction model, and the regional model MIj is trained. At the same time, the test set evaluation indicators MAE, RMSE, R 2 ; At the same time, based on the generalized meteorological data set and the corresponding regional data set, characteristic meteorological parameters are extracted from the meteorological parameters as the input features of the prediction model to train the generalized model MIJ; at the same time, the test set evaluation indicators MAE, RMSE, R 2 ; Then, the evaluation index of the regional model is compared with the evaluation index of the local model corresponding to each site, and the regional model or the local model is selected as the proposed prediction model; the evaluation index of the proposed prediction model is compared with the evaluation index of the generalized model as the final prediction model to obtain the ET0 prediction model; In the process of using the ET0 prediction model to predict the evapotranspiration of agricultural irrigation areas, the characteristic meteorological parameters obtained based on the meteorological parameters are first input into the ET0 prediction model for ET0 prediction.

2. The method for predicting evapotranspiration in an agricultural irrigation area according to claim 1, characterized in that: The characteristic meteorological parameters include Tmax, RH, Wind and LI.

3. A method for predicting evapotranspiration in an agricultural irrigation area according to claim 1 or 2, characterized in that: The machine learning models in S103 include XGboost, Random Forest RF, and Catboost.

4. A method for predicting evapotranspiration in agricultural irrigation areas, characterized in that: Use ET0 prediction model to predict evapotranspiration in agricultural irrigation areas; The ET0 prediction model is obtained by the following steps: A101. Obtain daily meteorological parameters of each meteorological station in the monitoring area within a period of time. The daily meteorological parameters include the year Year and the date day, and the corresponding meteorological parameters. The meteorological parameters include the maximum temperature Tmax and the minimum temperature Tmin, the average relative humidity RH, the wind speed Wind, the sunshine hours SH, the solar radiation flux RAD, and the light intensity LI; A102. Feature importance ranking and data division: The meteorological data of each meteorological station are recorded as station data sets; at the same time, based on the meteorological data of each meteorological station, the corresponding evapotranspiration ET0 values ​​are calculated using the FAO-56PM equation; The CatBoost algorithm is used to rank the meteorological parameters of each meteorological station. The top three meteorological parameters are selected and the meteorological stations with the same ranking are classified into one region. The obtained regions are recorded as region j, and the data sets of the same region are merged into a regional data set with feature correlation. At the same time, the top 4 meteorological parameters in the meteorological parameter importance ranking are taken as input features of the machine learning model; The top 4 meteorological parameters are recorded as characteristic meteorological parameters; A103. Based on region j, characteristic meteorological parameters are extracted from the corresponding regional data set as input features of the prediction model, and the output is the predicted ET0. The ET0 value calculated by the FAO-56PM equation is used as the label to train the prediction model corresponding to region j. Multiple machine learning models are trained for each region as prediction models, and the test set evaluation indicators of the model, such as MAE, RMSE, and R are recorded. 2 , and then select the optimal machine learning model as the ET0 prediction model corresponding to region j based on the evaluation index; In the process of using the ET0 prediction model to predict the evapotranspiration of agricultural irrigation areas, the characteristic meteorological parameters are first obtained based on the meteorological parameters, and the ET0 prediction model of the area where the characteristic meteorological parameters correspond to the station is input to perform ET0 prediction.

5. The method for predicting evapotranspiration in an agricultural irrigation area according to claim 4, characterized in that: The characteristic meteorological parameters include Tmax, RH, Wind and LI.

6. A method for predicting evapotranspiration in an agricultural irrigation area according to claim 4 or 5, characterized in that: The machine learning models in S103 include XGboost, Random Forest RF, and Catboost.

7. A method for predicting water level in an agricultural irrigation area, characterized in that: The following steps are involved: S201, obtaining historical water level data and evapotranspiration meteorological data, wherein the evapotranspiration meteorological data includes maximum temperature Tmax, minimum temperature Tmin, relative humidity RH and evapotranspiration ET0; S202, using fast Fourier transform to analyze the water level data, obtain the frequency result of the main peak of the water level data, and then determine the time series sliding window length N of the time series model based on the frequency result of the main peak of the water level data; S203, using VMD variational mode decomposition algorithm to perform VMD decomposition on the water level data to obtain 8 decomposition signals IMF1-IMF8; Based on IMF1~IMF8 and evapotranspiration meteorological data, the Pearson correlation coefficient method is used to determine the three IMFs in IMF1~IMF8 with high correlation coefficients with water level data, and the three IMFs, evapotranspiration meteorological data and water level data, a total of 8 features, constitute multivariate water level prediction data; S204, training water level prediction model: Based on 8 historical features, multivariate water level prediction data and a time series sliding window length N are formed, a feature matrix of (N, 8) is constructed as the input of a water level prediction model, and the predicted value of the future water level is used as the output of the water level prediction model to train the water level prediction model to obtain a trained water level prediction model; the neural network of the water level prediction model includes a one-dimensional convolutional neural network CNN layer and a GRU network set after the one-dimensional convolutional neural network CNN layer; S204. Predicting the future water level using the trained water level prediction model; when predicting the future water level, the future evapotranspiration data in the input data is the evapotranspiration data predicted by the agricultural irrigation area evapotranspiration prediction method according to claim 1 or 4.

8. A method for predicting water level in an agricultural irrigation area according to claim 7, characterized in that: The time series sliding window length N for determining the time series model is determined based on the period of the main peak 3 analyzed using fast Fourier transform.

9. The method for predicting water level in an agricultural irrigation area according to claim 7, characterized in that: The length N of the time series sliding window is 35 days.

10. The method for predicting water level in an agricultural irrigation area according to claim 7, characterized in that: The three IMFs among IMF1 to IMF8 that have high correlation coefficients with the water level data are IMF1, IMF6, and IMF7.

Images