CNN-LSTM combination model-based irrigation area underground water level burial depth prediction method

Abstract

The invention discloses an irrigation area underground water level burial depth prediction method based on a CNN-LSTM combination model, and the method comprises the steps: obtaining the data of the influence factors of the irrigation area underground water level burial depth, and carrying out the normalization processing; sorting the normalized data into a data set according to a time sequence, and dividing the data set into a training set and a test set; extracting data features of each influence factor of the underground water level burial depth; inputting the data features into a long short-term memory network to construct an underground water level burial depth prediction model, and finally obtaining the monthly underground water level burial depth after an output result is subjected to reverse normalization processing; and calculating a Nash efficiency coefficient NSE according to a prediction result, and performing parameter optimization on the CNN-LSTM combined prediction model to obtain an appropriate prediction model. According to the model, local and global information can be considered at the same time, so that the prediction result is higher in precision and stable; for an overfitting phenomenon easily occurring in small sample data, a method of discarding part of neurons is adopted, so that the problem can be effectively solved, and a relatively good prediction result can be obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of groundwater level depth prediction, and in particular to a groundwater level depth prediction method based on a CNN-LSTM combined model, which is suitable for irrigation areas with seasonal changes in groundwater level depth and control requirements. Background Art

[0002] Groundwater depth in irrigation areas influences crop growth, yield, and soil water and nutrient absorption. Appropriate groundwater depth can effectively improve farmland water and soil environments, promote crop growth, and increase crop yields. Particularly in my country's arid northwest region, excessively high groundwater levels can exacerbate secondary soil salinization, while excessively low levels can adversely impact terrestrial ecosystems such as farmland crops, natural grasslands, and woodlands. Furthermore, groundwater depth provides an important reference for groundwater extraction in irrigation areas. Therefore, accurate and reliable groundwater depth prediction plays a vital role in agricultural production, ecology, and water resource management in irrigation areas, and is crucial for the sustainable development of agriculture in these areas.

[0003] Currently, researchers often use regression analysis, time series analysis, neural networks, and other methods to analyze regional groundwater depth. However, these methods often suffer from issues such as inaccurate or inapplicable predictions. Further exploration and innovation are needed to compare and validate different methods to find groundwater depth predictions suitable for a specific region. To identify a suitable method for predicting groundwater depth in irrigation districts and improve the accuracy and stability of predictions, this paper proposes a groundwater depth prediction method for irrigation districts based on a CNN-LSTM combined model. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method for predicting the depth of groundwater level in irrigation areas, particularly based on a CNN-LSTM combined model, which is suitable for areas with seasonal changes in groundwater level depth and control requirements.

[0005] Technical solution: The method for predicting the groundwater depth in an irrigation area based on the CNN-LSTM combined model of the present invention comprises the following steps:

[0006] (1) Obtain data on factors affecting the depth of groundwater level in the irrigation area and perform normalization processing;

[0007] (2) Arrange the normalized data into a dataset in chronological order and divide the dataset into a training set and a test set;

[0008] (3) Input the training set data into the convolutional neural network to extract the data characteristics of various factors affecting the groundwater depth;

[0009] (4) The extracted data features are input into the long short-term memory network to construct a groundwater level prediction model, and the output results are denormalized to finally obtain the monthly groundwater level depth;

[0010] (5) The Nash efficiency coefficient NSE is calculated based on the prediction results, and the parameters of the CNN-LSTM combined prediction model are optimized to obtain a suitable irrigation area groundwater level depth prediction model.

[0011] Furthermore, the factors affecting the depth of the groundwater level in step (1) include monthly rainfall, Yellow River water diversion, evaporation, drainage, and the depth of the groundwater level in the previous month.

[0012] Furthermore, the calculation formula for data normalization in step (1) is:

[0013]

[0014] Among them, x max is the maximum value of the input sequence; x min is the minimum value of the input sequence; x and x′ are the data before and after normalization, respectively.

[0015] Furthermore, in step (3), the convolutional neural network extracts data features through the convolution layer, and the calculation of the convolution process is expressed as:

[0016] Y(i, j) = σ([ω]*[X i，j ]+b)

[0017]

[0018] Among them, Y(i, j) is the feature data output by CNN, σ is the tanh activation function, * is the convolution symbol, [ω] is the two-dimensional convolution kernel, [X i，j ] is the groundwater level in the convolution area, m is the number of rows where the data is located, n is the number of columns where the data is located, l is the convolution kernel moving step, and b is the bias parameter.

[0019] Furthermore, the groundwater level prediction model constructed in the long short-term memory network in step (4) adds a cell state component, and the updating process is regulated by three main gates: input gate, forget gate, and output gate.

[0020] Furthermore, in step (4), the convolutional neural network inputs the extracted data features into the long short-term memory network to construct a groundwater level prediction model, and the process is as follows:

[0021] Forget gate: f t =σ[W t ·(h t-1 , x t )+bf ]

[0022] Among them, f t is the output of the forget gate, σ is the sigmoid activation function, x t is the input information at time t, h t-1 is the output of the hidden layer at time t-1, W t is the learning weight of the forget gate, b f is the bias parameter of the forget gate;

[0023] Input gate: i t =σ[W i ·(h t-1 , x t )+b i ]

[0024] Among them, i t is the output of the input gate, W i is the learning weight of the input gate, b i is the bias parameter of the input gate;

[0025] Memory cell status

[0026] Among them, tanh is the activation function, W c is the learning weight of the memory cell, b c is the bias parameter of the memory cell;

[0027] The memory cell state update formula is:

[0028]

[0029] Output gate: o t =σ[W o ·(h t-1 , x t )+b o ]

[0030] Among them, t is the output of the output gate, W o is the learning weight of the output gate, b o is the bias parameter of the output gate.

[0031] Furthermore, the groundwater level prediction model constructed in the long short-term memory network discards some neurons.

[0032] Furthermore, the calculation method of the Nash efficiency coefficient NSE is:

[0033]

[0034] Among them, O iis the measured value of groundwater depth; O′ i is the measured average value of groundwater depth; P i is the predicted value of the CNN-LSTM model; n is the number of observations.

[0035] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: the convolutional neural network in the present invention can extract local features of data through convolution operations, and the long short-term memory network can capture long-term dependencies in sequence data, realize feature abstraction and high-level temporal pattern learning. The combination of the two enables the model to consider local and global information at the same time, making the prediction results more accurate and stable; the long short-term memory network model contains more parameters, and for small sample data prediction models, overfitting is prone to occur. Therefore, the method of discarding some neurons can effectively solve this problem and obtain better prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the method of the present invention;

[0037] Figure 2 Schematic diagram of CNN structure;

[0038] Figure 3 Schematic diagram of LSTM structure;

[0039] Figure 4 Comparison chart of NSE values ​​predicted 10 times by CNN-LSTM combination model and LSTM model;

[0040] Figure 5 This is a comparison chart between the predicted value of the CNN-LSTM combined model and the actual groundwater level depth. DETAILED DESCRIPTION

[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0042] A certain irrigation district belongs to the Northwest Large Irrigation District, with a total land area of ​​1.2×10 6 hm 2The 2012 irrigation district of Sichuan Province comprises five irrigation districts: District I, District II, District III, District IV, and District V, from west to east. Water resources in the irrigation district are primarily composed of three components: Yellow River diversion water, precipitation, and groundwater. Taking District III of this irrigation district as an example, based on statistical data from 2000 to 2012, this study analyzed the main factors influencing the groundwater depth in District III, including rainfall, Yellow River diversion water, evaporation, drainage, and the previous month's groundwater depth. Therefore, the model input parameters are monthly rainfall, Yellow River diversion water, evaporation, drainage, and the previous month's groundwater depth, and the output parameter is the actual monthly groundwater depth. Groundwater levels in District III were predicted using both an LSTM model and a CNN-LSTM combined model. The prediction results were compared and analyzed, providing a reference for selecting groundwater depth prediction methods in irrigation districts.

[0043] like Figure 1 As shown, the present invention includes the following basic steps:

[0044] (1) Obtain data on various factors affecting the groundwater depth in the irrigation area, including rainfall, Yellow River water diversion, evaporation, drainage, and the groundwater depth in the previous month, and normalize the above data;

[0045] Based on the statistical data of the irrigation district from 2000 to 2012, the main factors affecting the groundwater depth in the III irrigation area were analyzed, including rainfall, Yellow River water diversion, evaporation, drainage, and the groundwater depth of the previous month. The data normalization calculation formula is:

[0046]

[0047] Among them, x max is the maximum value of the input sequence; x min is the minimum value of the input sequence; x, x′ are the data before and after normalization respectively.

[0048] (2) The normalized data are organized into a dataset in chronological order, and the dataset is divided into a training set and a test set;

[0049] (3) The training set data is input into the convolutional neural network to extract the data features of various factors affecting the groundwater depth. The convolutional neural network is a deep learning model specially designed to process data with a grid structure. It consists of an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer. Figure 2 shown.

[0050] Here we further explain that the convolution layer completes the data feature extraction of various factors affecting the groundwater level depth through convolution operation. The calculation formula of the convolution process is:

[0051] Y(i, j) = σ([ω]*[Xi，j ]+b)

[0052]

[0053] Among them, Y(i, j) is the feature data output by CNN, σ is the tanh activation function, * is the convolution symbol, [ω] is the two-dimensional convolution kernel, [X i，j ] is the groundwater level in the convolution area, m is the number of rows where the data is located, n is the number of columns where the data is located, l is the convolution kernel moving step, and b is the bias parameter;

[0054] (4) The extracted data features are input into the long short-term memory network to construct a groundwater level prediction model, and the output results are denormalized to finally obtain the monthly groundwater level depth. Among them, the long short-term memory network is a derivative form of the recurrent neural network (RNN), which is specifically optimized for the gradient vanishing and gradient explosion problems that occur when RNN processes long sequences. LSTM introduces a unique gating mechanism, which enables the network to more effectively process and memorize long-term dependency information in the sequence, thereby improving its performance in time series data analysis. Compared with RNN, LSTM adds the "cell state" component, and its update process is regulated by three main gates, namely the input gate, the forget gate, and the output gate. Figure 3 shown.

[0055] The "cell state" update process is mainly regulated by the three gates mentioned above. CNN inputs the extracted data features into the long short-term memory network to build a groundwater level prediction model. The process is as follows:

[0056] Forget gate: f t =σ[W t ·(h t-1 , x t )+b f ]

[0057] Where: f t is the output of the forget gate, σ is the sigmoid activation function, x t is the input information at time t, h t-1 is the output of the hidden layer at time t-1, W t is the learning weight of the forget gate, b f is the bias parameter of the forget gate.

[0058] Input gate: i t =σ[W i ·(h t-1 , x t )+b i ]

[0059] Where: i tis the output of the input gate, W i is the learning weight of the input gate, b i is the bias parameter of the input gate.

[0060] Memory cell status

[0061] Where: tanh is the activation function, W c is the learning weight of the memory cell, b c is the bias parameter of the memory cell.

[0062] The memory cell state update formula is:

[0063]

[0064] Output gate: o t =σ[W o ·(h t-1 , x t )+b o ]

[0065] Where: t is the output of the output gate, W o is the learning weight of the output gate, b o is the bias parameter of the output gate.

[0066] Since the LSTM prediction model involves many parameters and the amount of data on various factors affecting the depth of groundwater level in the irrigation area is small, overfitting is prone to occur. The present invention adopts a method of discarding some neurons to effectively solve this problem.

[0067] (5) The Nash efficiency coefficient NSE was calculated based on the prediction results, and the parameters of the CNN-LSTM prediction model were optimized to obtain a suitable irrigation area groundwater depth prediction model. The Nash efficiency coefficient NSE was used to reflect the prediction accuracy and stability of the model. The closer the NSE is to 1, the better the prediction effect. In order to analyze the stability of the prediction model, the model was run 10 times independently to observe the changes in the NSE value. The formula of the Nash efficiency coefficient NSE is:

[0068]

[0069] Where: O i is the measured value of groundwater depth; O′ i is the measured average value of groundwater depth; P i is the predicted value of the CNN-LSTM model, and n is the number of observations.

[0070] The above CNN-LSTM combined model and LSTM model were used to predict the groundwater depth of irrigation area III. The groundwater depth of the irrigation area was predicted for 12 months in 2012 by Matlab software programming. The input variables are the five selected variables, and the output variable is the actual groundwater depth per month. The results are shown in Tables 1 and 2. Figure 4 The following is a comparison of the NSE values ​​predicted by the CNN-LSTM combination model and the LSTM model for 10 times; Figure 5 The figure shows a comparison between the predicted value of the CNN-LSTM combined model and the actual groundwater level depth.

[0071] Table 1 Prediction results of the CNN-LSTM combined model (10 independent predictions per run)

[0072]

[0073] Table 2 LSTM model prediction results (10 independent predictions per run)

[0074]

[0075]

[0076] The results show that the NSE values ​​of the CNN-LSTM combined model are stable and more accurate than those of the LSTM model, with most values ​​close to 0.8, while the LSTM model is mostly close to 0.78, indicating that the combined improvement is effective. Comparing the predicted values ​​of the CNN-LSTM combined model with the actual monthly groundwater levels, the predicted trends are consistent, and the prediction results are relatively accurate. This shows that the CNN-LSTM combined neural network model proposed in this paper can be used as a method for predicting the depth of groundwater levels in irrigation areas.

Claims

1. A method for predicting groundwater depth in irrigation areas based on a CNN-LSTM combined model, characterized in that: The steps include: (1) Obtain data on factors affecting the depth of groundwater level in the irrigation area and perform normalization processing; (2) Arrange the normalized data into a dataset in chronological order and divide the dataset into a training set and a test set; (3) Input the training set data into the convolutional neural network to extract the data characteristics of various factors affecting the groundwater depth; (4) The extracted data features are input into the long short-term memory network to construct a groundwater level prediction model, and the output results are denormalized to finally obtain the monthly groundwater level depth; (5) The Nash efficiency coefficient NSE is calculated based on the prediction results, and the parameters of the CNN-LSTM combined prediction model are optimized to obtain a suitable irrigation area groundwater level depth prediction model.

2. The method for predicting groundwater depth in irrigation areas based on the CNN-LSTM combined model according to claim 1 is characterized in that: The factors affecting the groundwater level depth in step (1) include the monthly rainfall, Yellow River water diversion, evaporation, drainage, and the groundwater level depth of the previous month.

3. The method for predicting groundwater depth in irrigation areas based on the CNN-LSTM combined model according to claim 1 is characterized in that: The calculation formula for data normalization in step (1) is: Among them, x max is the maximum value of the input sequence; x min is the minimum value of the input sequence; x and x′ are the data before and after normalization, respectively.

4. The method for predicting groundwater depth in irrigation areas based on the CNN-LSTM combined model according to claim 1, characterized in that: In step (3), the convolutional neural network extracts data features through the convolution layer, and the calculation of the convolution process is expressed as: Y(i,j)=σ([ω]*[X i，j ]+b) Among them, Y(i, j) is the feature data output by CNN, σ is the tanh activation function, * is the convolution symbol, [ω] is the two-dimensional convolution kernel, [X i，j ] is the groundwater level in the convolution area, m is the number of rows where the data is located, n is the number of columns where the data is located, l is the convolution kernel moving step, and b is the bias parameter.

5. The method for predicting groundwater depth in irrigation areas based on the CNN-LSTM combined model according to claim 1, characterized in that: The groundwater level prediction model constructed in the long short-term memory network in step (4) adds a cell state component, and the updating process is regulated by three main gates: input gate, forget gate, and output gate.

6. The method for predicting groundwater depth in irrigation areas based on the CNN-LSTM combined model according to claim 1, characterized in that: In step (4), the convolutional neural network inputs the extracted data features into the long short-term memory network to construct a groundwater level prediction model. The process is as follows: Forget gate: f t =σ[W t ·(h t-1 , x t )+b f ] Among them, f t is the output of the forget gate, σ is the sigmoid activation function, x t is the input information at time t, h t-1 is the output of the hidden layer at time t-1, W t is the learning weight of the forget gate, b f is the bias parameter of the forget gate; Input gate: i t =σ[W i ·(h t-1 , x t )+b i ] Among them, i t is the output of the input gate, W i is the learning weight of the input gate, b i is the bias parameter of the input gate; Memory cell status Among them, tanh is the activation function, W c is the learning weight of the memory cell, b c is the bias parameter of the memory cell; The memory cell state update formula is: Output gate: o t =σ[W o ·(h t-1 , x t )+b o ] Among them, t is the output of the output gate, W o is the learning weight of the output gate, b o is the bias parameter of the output gate.

7. The method for predicting groundwater depth in irrigation areas based on the CNN-LSTM combined model according to claim 6, characterized in that: The groundwater level prediction model constructed in the long short-term memory network discards some neurons.

8. The method for predicting groundwater depth in irrigation areas based on the CNN-LSTM combined model according to claim 1, characterized in that: The calculation method of the Nash efficiency coefficient NSE is: Among them, O i is the measured value of groundwater depth; O′ i is the measured average value of groundwater depth; P i is the predicted value of the CNN-LSTM model; n is the number of observations.

Images