Method for predicting water demand of irrigated area

Abstract

The invention provides an irrigation area water demand prediction method. The method comprises the following steps: acquiring historical meteorological characteristics and soil moisture content data in an irrigation area, and preprocessing the historical meteorological characteristics and soil moisture content data in the irrigation area to obtain preprocessed effective data; obtaining a corrected standard evapotranspiration value based on the preprocessed effective data and a standard evapotranspiration value formula; building a water demand prediction model, and training the water demand prediction model through the preprocessed effective data and the corrected standard evapotranspiration value; predicting through the trained water demand prediction model to obtain a predicted evapotranspiration value; and calculating the water demand of the irrigation area according to the predicted evapotranspiration value. According to the invention, real-time monitoring and prediction of the water demand of a large-scale irrigation area are realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent analysis of irrigation districts, and specifically relates to a method for predicting water demand in irrigation districts. Background Technology

[0002] In the field of agricultural water resource management, accurate prediction of crop water demand in irrigation areas is crucial for achieving efficient irrigation, ensuring agricultural production, and rationally allocating water resources. With the advancement of agricultural modernization, traditional experience-based irrigation methods are no longer sufficient to meet the demands of modern agriculture for refined water resource management, prompting researchers to continuously explore more scientific and accurate methods for predicting water demand.

[0003] Early research focused primarily on estimating crop water requirements through simple statistical analysis of meteorological data. For example, long-term monitoring of meteorological elements such as temperature, humidity, wind speed, and sunshine duration was used to establish correlation models between these parameters and crop water requirements using linear regression or empirical formulas. However, this approach oversimplifies the complex relationship between crop growth and the environment, neglecting soil characteristics, crop variety differences, and the spatiotemporal variability of meteorological conditions. This resulted in low prediction accuracy and an inability to effectively guide actual irrigation decisions.

[0004] In recent years, the rise of machine learning technology has brought new opportunities for crop water requirement prediction. Some studies have used machine learning algorithms, such as support vector machines (SVM) and decision trees, to build predictive models using historical meteorological data and crop growth stage data. These models have improved prediction accuracy to some extent, but still have limitations. On the one hand, the models have high requirements for the amount and quality of data, and are prone to overfitting or underfitting in small data scenarios; on the other hand, they have poor adaptability to different regions and different crop varieties, and lack generalization ability.

[0005] In practical applications, forecasting methods based on weather forecast information have also received widespread attention. These methods estimate water demand by acquiring weather forecast data and combining it with crop coefficients and water balance equations. However, weather forecast data has limitations in spatial resolution and accuracy, particularly in its insufficient description of field microclimates, leading to significant discrepancies between predicted results and actual field water demand.

[0006] In addition, some studies have focused on predicting evapotranspiration and attempting to indirectly obtain crop water demand information through the relationship between evapotranspiration and water requirement. For example, the PM formula is used to calculate the reference crop evapotranspiration value, and the actual evapotranspiration is measured using the eddy covariance method. However, in the process of converting evapotranspiration into crop water requirement, the determination of the crop coefficient is complex and affected by various factors, resulting in unstable and inaccurate water requirement predictions.

[0007] In summary, existing methods for predicting crop water requirements still need further improvement in terms of accuracy, universality, and practicality. There is an urgent need for an intelligent analysis method that can comprehensively consider multiple factors, adapt to different irrigation district environments and crop types, and has high prediction accuracy to meet the needs of modern agricultural irrigation management. Summary of the Invention

[0008] To address the above problems, this invention provides a method for predicting water demand in irrigation areas.

[0009] The technical solution of the present invention is as follows: A method for predicting water demand in an irrigation district, the method comprising: Obtain historical meteorological characteristics and soil moisture data within the irrigation area; The historical meteorological characteristics and soil moisture data of the irrigation area are preprocessed to obtain preprocessed effective data; The corrected standard evaporation value is obtained based on the preprocessed effective data and the standard evaporation value formula; A water demand prediction model is built, and the water demand prediction model is trained using the preprocessed effective data and the corrected standard evapotranspiration value. Predicted evapotranspiration values ​​are obtained by using a trained water demand prediction model. The water demand of the irrigation area is calculated based on the predicted evapotranspiration value.

[0010] Furthermore, the historical meteorological data includes daily average temperature, daily maximum temperature, daily minimum temperature, daily average wind speed at 2m above ground, daily average humidity, and sunshine duration; the soil moisture data includes soil temperature and humidity at 10cm depth, 20cm depth, 30cm depth, and 40cm depth.

[0011] Furthermore, the preprocessing of the historical meteorological characteristics and soil moisture data includes: Using the moving window method, the average of 7 days of window data is used to perform linear interpolation at the missing locations; Using percentile and time series analysis methods, values ​​below the 5th percentile, above the 95th percentile, and those with a difference exceeding 3 times the standard deviation were marked as outliers, and outliers and dirty data points were deleted.

[0012] Furthermore, the process of obtaining the corrected standard evaporation value based on the preprocessed valid data and the standard evaporation value formula is as follows: First, calculate the standard evaporation rate using the standard evaporation value formula; The water stress factor of the soil in the tank area was calculated using the water stress factor formula. The corrected standard evapotranspiration value is obtained by multiplying the standard evapotranspiration value by the moisture stress factor of the soil in the tank area.

[0013] Furthermore, the water demand prediction model is based on a two-layer structure of Stacking ensemble learning, including a primary learner and a secondary learner, wherein; The primary learner is implemented using LSTM and Transformer time series to extract time series features of various meteorological characteristics and standard evapotranspiration values; The secondary learner uses ridge regression to weight and fuse the prediction results of the primary learner to generate the final prediction result.

[0014] Furthermore, the loss function of the ridge regression method is: In the formula, For loss function, Represents the actual observed values ​​of the dependent variable. Let W be the characteristic variable and W be the regression coefficient. This is the penalty coefficient.

[0015] Furthermore, the parameter settings for training the water demand prediction model include: The primary learner takes 13 feature inputs, including: 6 meteorological factors required for the standard evapotranspiration formula calculation and historical water demand data for the past seven days; the output feature has a dimension of 1. The number of hidden layer units in the primary learner is set to 128, and the number of layers is set to 2. The time step is set to 7; The Adam optimizer was used to optimize the parameters, with the initial learning rate set to 0.001. The model training process involves 100 iterations. Set the batch size to 32.

[0016] Furthermore, the input features are normalized to map the data range to the [0, 1] interval; The output features are also normalized.

[0017] Furthermore, the parameter settings for training the water demand prediction model also include: Add a learning rate adjuster, StepLR, which periodically adjusts the learning rate during training by a set step size. The adjustment formula is as follows: in, The adjusted learning rate, The initial learning rate, This is the learning rate decay factor. This represents the current number of training iterations. The step size for adjusting the learning rate.

[0018] Furthermore, the calculation of the irrigation district's water demand based on the predicted evapotranspiration value specifically involves: Based on the predicted evapotranspiration value combined with the crop coefficient method, the crop water requirement is calculated according to the crop type determined in the irrigation area. The water demand of the irrigation area is calculated based on the actual crop planting area and crop water requirements in the irrigation area.

[0019] Compared with the prior art, the present invention has the following advantages: This invention innovatively combines LSTM and Transformer deep learning models, employing a stacking ensemble learning approach. LSTM effectively captures both long-term and short-term dependencies in water demand, uncovering dynamic features from time-series data; Transformer utilizes a self-attention mechanism to acquire global features and long-term correlations in meteorological data, enhancing its ability to model complex time series. The complementary strengths of these two models enable the model to comprehensively capture the spatiotemporal characteristics of water demand changes, offering a significant advantage over single models in processing complex meteorological data and effectively overcoming their limitations.

[0020] This invention corrects the standard evapotranspiration value calculated by the FAO56-PM formula by introducing a soil moisture stress factor, thus no longer relying solely on meteorological data. This method fully considers the dynamic changes in soil moisture, making the evapotranspiration value calculation more closely reflect the actual conditions of the irrigation area. It solves the problem of large prediction errors caused by differences between meteorological and field microclimates in traditional methods, improves the reliability of data input, and thus makes the water demand prediction results more consistent with the actual planting environment and water requirements of the irrigation area. Attached Figure Description

[0021] The accompanying drawings illustrate various embodiments generally by way of example rather than limitation, and are used, together with the specification and claims, to explain embodiments of the invention. Where appropriate, the same reference numerals are used in all drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the apparatus or method.

[0022] Figure 1 A flowchart illustrating the prediction process for the standard evapotranspiration value of this invention is shown. Figure 2 A flowchart illustrating the predicted water demand calculation of the present invention is shown; Figure 3 The prediction effect diagram of the Stacking model of the present invention is shown; Figure 4 The diagram showing the water demand prediction effect of the present invention is illustrated. Detailed Implementation

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] like Figures 1-4 As shown, this embodiment of the invention provides a method for predicting water demand in irrigation areas; the method includes the following steps: (1) acquiring historical meteorological characteristics and soil moisture data; (2) data preprocessing; (3) correcting the FAO56-PM formula; (4) building a Stacking ensemble learning model; (5) setting and optimizing model parameters; and (6) predicting results and calculating water demand. (1) Acquisition of historical meteorological characteristics and soil moisture data Historical meteorological data and soil moisture data are acquired using sensors and meteorological stations within the irrigation area. This data is then wirelessly transmitted to a data terminal. Meteorological data, depending on calculation needs, includes: daily average temperature, daily maximum temperature, daily minimum temperature, daily average wind speed at 2m above ground, daily average humidity, and sunshine duration. Soil moisture data includes: soil temperature and humidity at 10cm depth, 20cm depth, 30cm depth, and 40cm depth.

[0025] (2) Data preprocessing For missing values, a moving window method is used to linearly interpolate the average of the 7-day window of data to the missing location. Percentile and time series analysis methods are used to mark values ​​below the 5th percentile, above the 95th percentile, and those with differences exceeding three times the standard deviation as outliers, and these outliers and dirty data points are then removed.

[0026] (3) FAO56-PM Formula Correction The standard evapotranspiration formula (FAO 56-PM) recommended by the Food and Agriculture Organization of the United Nations is as follows: (1) In the formula For reference standard evaporation, G represents net radiation to the crop surface, T represents soil heat flux, and T represents the average daily air temperature at a height of 2m. The wind speed at a height of 2 meters. It is the saturated vapor pressure. Actual vapor pressure The slope of the vapor pressure curve. is the humidity constant.

[0027] The formula calculation results are corrected by calculating the soil moisture stress factor. The specific calculation steps are as follows: Determine the soil type of the experimental area, taking sandy loam as an example, and calculate the field water holding capacity (…). The percentage is generally 10%-12%, with permanent wilting points ( The soil moisture stress factor is generally between 4% and 6%. It can be calculated using formula (2): (2) Where S represents soil moisture. Since 20cm is the main root growth depth for most crops, a soil depth of 20cm is selected as the correction calculation parameter. The standard evapotranspiration value calculated using the FAO56-PM formula is multiplied by the soil moisture stress factor to obtain the corrected standard evapotranspiration value, providing data for subsequent model training and water requirement calculation.

[0028] (4) Building a Stacking ensemble learning model To improve the accuracy of irrigation demand prediction, this invention designs a model based on Stacking ensemble learning, which combines the advantages of deep learning and traditional machine learning algorithms to fully explore the temporal characteristics and nonlinear relationships in historical meteorological features and soil moisture data.

[0029] a. Stacking Model Architecture This model employs a two-layer structure, using LSTM and Transformer time series as the primary learner to extract time series features of various meteorological characteristics and standard evapotranspiration values. The secondary learner uses ridge regression to weight and fuse the prediction results of the primary learner to generate the final prediction result.

[0030] b. LSTM model LSTM (Long Short-Term Memory) is a special type of recurrent neural network specifically designed to capture long-term and short-term dependencies in time-series data. By introducing forget gates, input gates, and output gates, it effectively solves the vanishing and exploding gradient problems found in traditional RNNs.

[0031] c. Transformer time series model The Transformer is a self-attention-based neural network widely used in natural language processing and time series analysis. Unlike traditional recurrent structures, the Transformer captures global temporal dependencies through parallel computation, making it particularly suitable for long-running series. The Transformer model primarily consists of two parts: an encoder and a decoder. Each encoder layer comprises two core components: a multi-head self-attention mechanism and a forward propagation network. The decoder employs a similar structure but adds a second attention mechanism to directly focus on the encoder's output. This design allows the Transformer to process all parts of the input data simultaneously, significantly improving processing speed and efficiency.

[0032] d. Ridge Return Ridge regression is a type of regression method, belonging to statistical methods. It is a least squares regression with L2 penalty. When the input variables are strongly correlated, fitting the data using least squares can lead to significant bias. Furthermore, multicollinearity exists among these variables; if this multicollinearity is not addressed, the research results will be highly biased. The ridge regression loss function is: In the formula Let X be the loss function, X be the feature variable, and W be the regression coefficient. As a penalty coefficient, by compressing the parameters of the explained variable, it can effectively solve the problem of multicollinearity among the explanatory variables and improve the robustness of the research results.

[0033] (5) Model parameter setting and tuning Parameter settings: The primary learner uniformly adopts 13 feature inputs. In addition to the 6 meteorological factors required for the FAO56-PM formula calculation, it also adds historical water demand data for the past seven days.

[0034] Table 1 Feature Input Parameters

[0035] To enhance the model's expressive power, the number of hidden units (hidden_size) in the primary learner is set to 128 to capture complex time-series patterns in the features. To improve the model's ability to learn deep temporal dependencies, both primary learners have a layer count (num_layers) of 2. The Transformer model's multi-head attention heads (num_heads) are set to 2. Output feature dimension (output_size): The model output is the predicted reference evapotranspiration (ET0-FAO56) for one day, therefore the output dimension is set to 1. Time step (look_back): The time step is set to 7, meaning the model uses input data from the past 7 days to predict the reference evapotranspiration for the next day. Optimizer and learning rate: The Adam optimizer is used for parameter optimization, with an initial learning rate of 0.001 to balance convergence speed and model performance. Loss function: Mean squared error (MSE) is used. in, For predicted values, is the actual value, and n is the number of samples.

[0036] Number of training iterations (num_epochs): The model is trained for 100 iterations to ensure convergence. Batch size (batch_size): The batch size is set to 32 to balance computational efficiency and gradient update stability.

[0037] Model tuning: All input features are normalized using MinMaxScaler, mapping the data range to the [0, 1] interval, as shown in the following formula: in, and These represent the minimum and maximum values ​​of the feature, respectively. The output target variable (ET0-FAO56) is normalized separately using MinMaxScaler to maintain consistency in numerical range. Both the training data and the prediction data from the primary learner are normalized.

[0038] Learning Rate Adjustment: To accelerate model convergence, a learning rate adjuster, StepLR, is added. This adjusts the learning rate periodically according to a set step size during training, helping the optimizer train the model more effectively. It reduces the learning rate in stages to prevent the model from oscillating or getting stuck in local optima in later training phases. Specifically, the learning rate is reduced to 0.1 times its original value every 20 epochs, with the update formula as follows: in The initial learning rate, = 0.1 is the learning rate decay factor.

[0039] (6) Result prediction and water demand calculation: a. Outcome prediction Two primary learners are trained separately, using mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (COP). The predictive performance of the model is evaluated.

[0040] MSE: MAE: : in, For predicted values, This is the actual value. Let n be the mean and n be the number of samples. After normalizing the predictions from the primary learner, ridge regression is used for final fitting to obtain the final prediction result.

[0041] Table 2 Model Evaluation

[0042] b. Water demand calculation Water requirements are calculated based on the crop type identified in the irrigation district using predicted evapotranspiration values ​​combined with the crop coefficient method recommended by the FAO. Crop coefficient method: in, Crop water requirements The crop coefficient is a measure of the crop growth cycle, generally divided into four stages: early stage, middle stage, peak stage, and late stage. This is the corrected standard evapotranspiration value predicted by the Stacking model.

[0043] Table 3 Partial Prediction Results

[0044] Furthermore, the total water requirement can be calculated based on the actual crop planting area in the irrigation district, using the following conversion formula: in Total water demand ( ), To predict water demand (mm / d), Area represents the planting area ( ).

[0045] This invention combines the advantages of LSTM and Transformer, effectively capturing the time-series characteristics of irrigation district water demand and the complex relationships among multiple variables through a Stacking ensemble learning architecture. By introducing ridge regression as a secondary learner to weight and optimize the prediction results from LSTM and Transformer, the prediction accuracy of the final model is further improved, overcoming the limitations of single models in handling complex meteorological data. Using time-series features (lagging water demand) as additional input, the model can more accurately reflect the dynamic trends of irrigation district water demand. The Stacking ensemble model effectively combines the long-term dependency capture capability of LSTM and the global feature extraction capability of Transformer, giving the model stronger generalization ability when facing different irrigation district conditions (such as climate differences and crop types), thus enhancing the model's robustness. The model has a simple structure and is easy to deploy into existing irrigation district management systems, enabling real-time monitoring and prediction of water demand in large-scale irrigation districts through automated calculations.

[0046] 1. The principle of improvement FAO56-PM Formula Correction Principle: The standard FAO56-PM formula uses only meteorological data for calculation. By introducing a soil moisture stress factor correction formula, the evapotranspiration calculation becomes closer to the actual conditions in the irrigation area. The corrected evapotranspiration value not only considers meteorological conditions but also incorporates dynamic changes in soil moisture, solving the problem of large prediction errors caused by differences between meteorological and field microclimates in traditional methods.

[0047] The ensemble learning principle of Stacking models: Stacking ensemble learning uses primary learners (LSTM and Transformer) to extract time-series features of water demand and model nonlinear relationships, transforming complex data patterns into simple feature representations. Secondary learners (ridge regression) further reduce the uncertainty of single-model predictions by weightedly fusing the predictions of the primary learners, thereby improving the accuracy of the final prediction.

[0048] 2. The Improvement Process The improvement process of this invention can be divided into the following three stages: Data correction and preprocessing: The standard evapotranspiration value was corrected using measured soil moisture data to ensure the quality of the input data. Meteorological and soil moisture data were normalized to avoid model bias due to differences in feature scales.

[0049] Elementary learner construction: LSTM model: By using forget gate, input gate and output gate mechanism, it captures the long-term and short-term dependencies of water demand and mines the dynamic features of time series data.

[0050] Transformer model: It uses a self-attention mechanism to capture the global features and correlations of meteorological data over long time spans, thereby enhancing the model's ability to model complex time series.

[0051] Secondary learner optimization: The output of the primary learner serves as the input to the secondary learner. The ridge regression model effectively reduces the impact of multicollinearity on the final prediction by introducing a regularization term.

[0052] 3. How can the improved methods achieve positive results? Improved data correction methods: After formula revision, standard evapotranspiration is dynamically coupled with soil moisture, improving the reliability of data input. The water demand prediction results are now more closely aligned with the actual planting environment and water requirements of the irrigated area.

[0053] Improve model architecture and learning methods: The LSTM model is used to deeply extract historical sequence features of water demand in irrigation districts, and combined with the global dependency modeling capability of the Transformer model, the spatiotemporal features of water demand changes are comprehensively captured. Through the ensemble learning strategy of the Stacking model, the advantages of primary learners are effectively integrated, compensating for the limitations of single models when dealing with complex data features, resulting in more stable prediction performance.

[0054] Improved model optimization strategy: A learning rate adjustment mechanism (StepLR) is introduced during model training. A high learning rate is maintained in the early stages of training to accelerate convergence, while the learning rate is gradually reduced in the later stages to enhance model stability. Dynamic monitoring of the validation set, combined with an early stopping strategy, avoids overfitting and improves the model's generalization ability.

[0055] 4. Specific effects of the improvement The revised data and optimized model architecture significantly reduced the water demand forecasting error, lowering the MAE to 0.29. The accuracy was improved to 0.87, making it suitable for long-term forecasting needs in irrigated areas. Multi-source data input and ensemble learning strategies enable the model to be applied to different irrigated area environments and crop types, significantly improving the applicability and robustness of the forecast results.

[0056] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting water demand in an irrigation area, characterized in that, The method includes: Obtain historical meteorological characteristics and soil moisture data within the irrigation area; The historical meteorological characteristics and soil moisture data of the irrigation area are preprocessed to obtain preprocessed effective data; The corrected standard evaporation value is obtained based on the preprocessed effective data and the standard evaporation value formula; A water demand prediction model is built, and the water demand prediction model is trained using the preprocessed effective data and the corrected standard evapotranspiration value. Predicted evapotranspiration values ​​are obtained by using a trained water demand prediction model. The water demand of the irrigation area is calculated based on the predicted evapotranspiration value.

2. The irrigation district water demand prediction method according to claim 1, characterized in that, The historical meteorological data includes daily average temperature, daily maximum temperature, daily minimum temperature, daily average wind speed at 2m above ground, daily average humidity, and sunshine duration; the soil moisture data includes soil temperature and humidity at 10cm depth, 20cm depth, 30cm depth, and 40cm depth.

3. The irrigation district water demand prediction method according to claim 1, characterized in that, Preprocessing of the historical meteorological characteristics and soil moisture data includes: Using the moving window method, the average of 7 days of window data is used to perform linear interpolation at the missing locations; Using percentile and time series analysis methods, values ​​below the 5th percentile, above the 95th percentile, and those with a difference exceeding 3 times the standard deviation were marked as outliers and then deleted.

4. The irrigation district water demand prediction method according to claim 1, characterized in that, The modified standard evaporation value obtained based on the preprocessed effective data and the standard evaporation value formula is specifically as follows: First, calculate the standard evaporation rate using the standard evaporation value formula; The water stress factor of the soil in the tank area was calculated using the water stress factor formula. The corrected standard evapotranspiration value is obtained by multiplying the standard evapotranspiration value by the moisture stress factor of the soil in the tank area.

5. The method for predicting irrigation area water demand according to claim 1, characterized in that, The water demand prediction model is based on a two-layer structure of Stacking ensemble learning, including a primary learner and a secondary learner, wherein; The primary learner is implemented using LSTM and Transformer time series to extract time series features of various meteorological characteristics and standard evapotranspiration values; The secondary learner uses ridge regression to weight and fuse the prediction results of the primary learner to generate the final prediction result.

6. The irrigation district water demand prediction method according to claim 5, characterized in that, The loss function of the ridge regression method is: In the formula, For loss function, Represents the actual observed values ​​of the dependent variable. Let W be the characteristic variable and W be the regression coefficient. This is the penalty coefficient.

7. The method for predicting irrigation district water demand according to claim 1, characterized in that, The parameter settings for training the water demand prediction model include: The primary learner uses 13 input features, including: 6 meteorological factors required for the standard evapotranspiration formula calculation and historical water demand data for the past seven days; the output feature has a dimension of 1. The number of hidden layer units in the primary learner is set to 128, and the number of layers is set to 2. The time step is set to 7; The Adam optimizer was used to optimize the parameters, with the initial learning rate set to 0.

001. The model training process involves 100 iterations. Set the batch size to 32.

8. The method for predicting irrigation district water demand according to claim 7, characterized in that, The input features are normalized to map the data range to the interval [0, 1]. The output features are also normalized.

9. The method for predicting irrigation district water demand according to claim 1, characterized in that, The parameter settings for training the water demand prediction model also include: Add a learning rate adjuster, StepLR, which periodically adjusts the learning rate during training by a set step size. The adjustment formula is as follows: in, The adjusted learning rate, The initial learning rate, This is the learning rate decay factor. This represents the current number of training iterations. The step size for adjusting the learning rate.

10. The method for predicting irrigation district water demand according to claim 1, characterized in that, The water demand of the irrigation area is calculated based on the predicted evapotranspiration value as follows: Based on the predicted evapotranspiration value combined with the crop coefficient method, the crop water requirement is calculated according to the crop type determined in the irrigation area. The water demand of the irrigation area is calculated based on the actual crop planting area and crop water requirements in the irrigation area.

Images