A saltwater tide prediction method based on LSTM-GRU-CNN integrated model

Abstract

The present invention discloses a saltwater tide prediction method based on LSTM-GRU-CNN integrated model, and the method comprises the following steps: step 1, constructing a comprehensive data set; step 2, data preprocessing; step 3, constructing and training an LSTM-GRU-CNN integrated model; step 4, model testing and evaluation; step 5, saltwater tide prediction. The saltwater tide prediction method based on LSTM-GRU-CNN integrated model of the present invention solves the technical problems of large data demand and low prediction accuracy in the current saltwater tide prediction field. Compared with the currently commonly used numerical simulation methods, the method of the present invention has small data demand and simple modeling process, can realize accurate prediction of saltwater tide, improves the accuracy and efficiency of saltwater tide prediction, and has wide applicability and effectiveness.

Description

Technical Field

[0001] The invention belongs to the technical field of saltwater tide prediction, and in particular relates to a saltwater tide prediction method based on an LSTM-GRU-CNN integrated model. Background Art

[0002] Salt tide refers to a natural hydrological phenomenon in which seawater flows back into freshwater estuaries or rivers, causing the water to become salty. It usually occurs in winter or dry seasons, and appears at the confluence of rivers and seas, such as the surrounding areas of the Yangtze River Delta and the Pearl River Delta. Salt tide disasters have become an important factor restricting the development and utilization of water resources and economic sustainable development. The current salt tide prediction methods in my country mainly include field observation and analysis, physical model experiments, and numerical simulation methods. However, the field observation and analysis method is limited by observation equipment and human resources, and it is difficult to achieve continuous and large-scale monitoring. The prediction accuracy is also affected by the subjective judgment of the observer; the physical model test method is often limited by the model scaling effect and experimental environment factors, and it is difficult to fully reproduce the real estuary environment; the numerical simulation method has the problem of low prediction accuracy, and most studies are still limited to two-dimensional mathematical models, and the application of three-dimensional mathematical models is not yet common; the use of deep learning methods is relatively small. Therefore, how to improve the prediction accuracy of salt tide is a technical problem that needs to be solved urgently. Summary of the invention

[0003] The object of the present invention is to provide a saltwater tide prediction method based on an LSTM-GRU-CNN integrated model to solve the above-mentioned technical problems.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] The present invention discloses a saltwater tide prediction method based on an LSTM-GRU-CNN integrated model, and the method comprises the following steps:

[0006] Step 1: Construct a comprehensive data set: Obtain the measured estuary salinity data, upstream flow data, and estuary tide data of multiple pumping stations in a certain period of time in the study area to construct a comprehensive data set;

[0007] Step 2: Data preprocessing: preprocess the data in the comprehensive data set, and divide the preprocessed data into training period data and test period data in proportion;

[0008] Step 3, constructing and training the LSTM-GRU-CNN integrated model: constructing the LSTM-GRU-CNN integrated model, using the training period data to train the model, setting the lag time steps and the prediction time steps, using the comprehensive data of the training period with the set lag time steps as the model input, using the measured estuary salinity data of each pumping station during the training period with the set prediction time steps as the model output, and training the model until the model training is completed;

[0009] Step 4, model testing and evaluation: the comprehensive data of the test period data is input into the trained integrated model, where the lag time steps and the prediction time steps are consistent with those during training, and then the model prediction results are obtained. The model prediction results are denormalized to obtain the predicted estuary salinity data, and the predicted estuary salinity data is compared with the measured estuary salinity data of each pumping station, and the evaluation index is used to determine whether the model has passed the test;

[0010] Step 5: Salt tide prediction: Use the tested model to predict the salinity of the estuary in the target period of the study area to perform the corresponding salt tide prediction work.

[0011] Furthermore, the preprocessing process in step 2 includes data merging, missing value filling, data normalization and format conversion;

[0012] The data is merged to generate a time series containing hourly time points according to the minimum and maximum values ​​of the series, and the original data and the time series are merged using left connection to ensure that all time points are included;

[0013] The missing value filling is to use the preceding filling method to complete the missing salinity, flow and tide level data, that is, to fill the missing value with the nearest non-missing value before the missing value, so as to obtain complete data;

[0014] The data normalization is to use the MinMaxScaler class in the scikit-learn library in Python to perform data normalization and scale the data feature values ​​to the interval [0,1];

[0015] The format conversion converts the original data into a format for supervised learning, that is, the data is divided into two parts: input features and output labels, and the corresponding relationship between them is maintained.

[0016] Furthermore, the division ratio of the training period data and the test period data in step 2 is 7:3.

[0017] Furthermore, the specific process of constructing the LSTM-GRU-CNN integrated model in step 3 is:

[0018] Step 31: Use the LSTM, Conv1D, MaxPooling1D, Flatten, GRU, Dense, and Dropout layers of the layers module in the Keras library in Python to build an LSTM model, a GRU model, and a CNN model, respectively, and set the initial parameters of the three models respectively;

[0019] Step 32: Build an input layer to receive data, and connect the output of the input layer to the input of the LSTM model, GRU model, and CNN model respectively;

[0020] Step 33: Use the Concatenate layer of the layers module in the Keras library to connect the outputs of the three models together and fuse the features they extracted;

[0021] Step 34: Add a Dense layer to the concatenated output, whose output size is the number of features multiplied by the number of output time steps, and the activation function uses ReLU to obtain the final stacked output;

[0022] Step 35: Use the Model class of the models module in the Keras library to build an LSTM-GRU-CNN integrated model. The input of the integrated model is the input layer built in step 32, and the output of the model is the stacked output obtained in step 34.

[0023] Furthermore, the specific process of training the model in step 3 until the model training is completed is:

[0024] (1) First, a common loss function is used. According to the prediction results of the initial model using the common loss function, weights are set for the multiple pumping stations described in step 1 to adjust the model's emphasis on different pumping stations;

[0025] (2) Define a weighted loss function and use the mean square logarithmic error to calculate the relative error between the predicted value and the true value of each pumping station; since the data of multiple pumping stations are included in the comprehensive data set, the data of the corresponding row and column positions of each pumping station are extracted through Python slicing operations;

[0026] (3) Calculate the loss of the corresponding pumping station according to the weight, add up the weighted sum of the losses of each pumping station, and get the total loss value as the return value of the function. After the training is completed, output the change curve of the weighted loss function and analyze the image. When the function converges, it means that the model training is completed.

[0027] Furthermore, the specific process of using the evaluation index in step 4 to determine whether the model has passed the test is as follows:

[0028] The evaluation indicators include: Nash efficiency coefficient NSE, precision, accuracy, recall rate and the ratio of the predicted and measured saltwater tide times of each pumping station;

[0029] The calculation formula of NSE is:

[0030]

[0031] Where: Y i is the measured value at the i-th moment; is the predicted value at the i-th moment; is the average value of all measured values; n is the total number of measured values;

[0032] The calculation formula of accuracy is:

[0033]

[0034] The calculation formula of accuracy is:

[0035]

[0036] The calculation formula of recall rate is:

[0037]

[0038] Where: TP is the number of samples correctly predicted as positive; TN is the number of samples correctly predicted as negative; FP is the number of samples incorrectly predicted as negative as positive; FN is the number of samples incorrectly predicted as positive as negative; positive refers to salty tide, negative refers to non-salty tide;

[0039] The calculation formula for the ratio of the predicted and measured saltwater tide frequency is as follows:

[0040]

[0041] Calculate each evaluation index of the model and determine whether each evaluation index reaches the qualified threshold. When each evaluation index reaches the set qualified threshold, it means that the model test has passed.

[0042] Furthermore, the qualified threshold of NSE is 0.6; the qualified thresholds of precision, accuracy and recall are all 0.7; and the qualified threshold interval of the ratio of the predicted and measured saltwater tide times is [0.8, 1.2].

[0043] The beneficial effects of the present invention are as follows: the saltwater tide prediction method based on the LSTM-GRU-CNN integrated model of the present invention solves the technical problems existing in the current saltwater tide prediction field, such as large data demand and low prediction accuracy. Compared with the currently commonly used numerical simulation methods, the method of the present invention has small data demand and simple modeling process, can realize accurate prediction of saltwater tide, improves the accuracy and efficiency of saltwater tide prediction, and has wide applicability and effectiveness.

[0044] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the method of the present invention;

[0046] Figure 2Schematic diagram of the process of building the LSTM-GRU-CNN integrated model. DETAILED DESCRIPTION

[0047] The present invention discloses a saltwater tide prediction method based on LSTM-GRU-CNN integrated model, such as Figure 1 As shown, the method comprises the following steps:

[0048] Step 1: Construct a comprehensive dataset: Obtain the measured estuary salinity data, upstream flow data, and estuary tide data from multiple pumping stations in the study area during a certain period of time to construct a comprehensive dataset.

[0049] Step 2: Data preprocessing: Preprocess the data in the comprehensive data set. The preprocessing process includes data merging, missing value filling, data normalization, and format conversion. The preprocessed data is divided into training period data and test period data according to a certain ratio.

[0050] The specific steps include:

[0051] Step 21: Generate a time series containing hourly time points based on the minimum and maximum values ​​of the series, and use left join to merge the original data and the time series to ensure that all time points are included;

[0052] Step 22: Use the preceding filling method to complete the missing salinity, flow and tide data, that is, fill the missing value with the nearest non-missing value before the missing value, and then obtain complete data;

[0053] Step 23: Use the MinMaxScaler class of the scikit-learn library in Python to normalize the data and scale the data feature values ​​to the interval [0,1];

[0054] Step 24: Define a function to convert the original data into the format of supervised learning, that is, divide the data into two parts: input features and output labels, and maintain the corresponding relationship between them to facilitate the subsequent training and prediction of the supervised learning model;

[0055] Step 25: Divide the processed data into training data and test data in a certain ratio. The specific ratio should be determined according to actual needs and data characteristics. The commonly used ratio of training data to test data is 7:3.

[0056] Step 3, build an LSTM-GRU-CNN integrated model and train it: build an LSTM-GRU-CNN integrated model, use the training period data to train the model, set the lag time steps and the prediction time steps, use the comprehensive data of the training period with the set lag time steps as the model input, and use the measured estuary salinity data of each pumping station in the training period with the set prediction time steps as the model output, and train the model until the model training is completed.

[0057] like Figure 2 As shown, the specific steps include:

[0058] Step 31: Use the LSTM, Conv1D, MaxPooling1D, Flatten, GRU, Dense, and Dropout layers of the layers module in the Keras library in Python to build an LSTM model, a GRU model, and a CNN model, respectively, and set the initial parameters of the three models respectively;

[0059] Step 32: Build an input layer to receive data, and connect the output of the input layer to the input of the LSTM model, GRU model, and CNN model respectively;

[0060] Step 33: Use the Concatenate layer of the layers module in the Keras library to connect the outputs of the three models together and fuse the features they extracted;

[0061] Step 34: Add a Dense layer to the concatenated output, with an output size of the number of features multiplied by the number of output time steps, and a ReLU activation function to get the final stacked output; this shows that the prediction for each time step is a vector determined by the number of features, and the number of times this vector is repeated over the entire time series is equal to the number of output time steps;

[0062] Step 35: Use the Model class of the models module in the Keras library to build an LSTM-GRU-CNN integrated model. The input of the integrated model is the input layer built in step 32, and the output of the model is the stacked output obtained in step 34.

[0063] Step 36: Use the training period data to train the integrated model, set the lag time step and the prediction time step, use the comprehensive data of the training period with the set lag time step as the model input, and use the measured estuary salinity data of each pumping station during the training period with the set prediction time step as the model output; during training, first use the commonly used ordinary loss function, and then use the customized weighted loss function based on the result analysis to improve the forecast accuracy of all stations.

[0064] The role of the loss function is to measure the difference or degree of error between the model's predicted value and the true value. The loss function is the objective function of model optimization and is used to guide the update of model parameters so that the model's predicted value is as close to the true value as possible. During the training process, after each training batch, the model will calculate the current loss function value (i.e., the difference between the predicted and true values), find the derivative of the function, and optimize the model's parameters along the direction of the derivative (i.e., the direction of minimizing the loss value). After each parameter update, the value of the loss function is calculated to check the training effect.

[0065] The specific process includes the following:

[0066] (1) first use a common loss function, and according to the analysis of the prediction results of the initial model using the common loss function, set targeted weights for the multiple pumping stations described in step 1 to adjust the model's emphasis on different pumping stations;

[0067] (2) Define a weighted loss function and use the mean square logarithmic error (MSLE) to calculate the relative error between the predicted value and the true value of each pumping station. Since the data of multiple pumping stations are included in the comprehensive data set, it is necessary to use Python's slicing operation to extract the data of the corresponding row and column positions of each pumping station.

[0068] (3) Calculate the loss of the corresponding pumping station according to the weight, add up the weighted losses of each pumping station, and get the total loss value as the return value of the function. After the training is completed, output the change curve of the weighted loss function and analyze the image. When the function tends to be stable and no longer changes significantly (that is, the function can converge), it means that the model training is completed.

[0069] Step 4, model testing and evaluation: The comprehensive data of the test period data is input into the trained integrated model, where the lag time steps and the prediction time steps are consistent with those during training, and then the model prediction results are obtained. The model prediction results are denormalized to obtain the predicted estuary salinity data, and the predicted estuary salinity data are compared with the measured estuary salinity data of each pumping station. The evaluation indicators are used to evaluate the reliability of the model, that is, to determine whether the model has passed the test.

[0070] Among them, the selected evaluation indicators are Nash efficiency coefficient NSE, precision, accuracy, recall rate and the ratio of predicted and measured saltwater tide times at each pumping station.

[0071] The calculation formula of NSE is:

[0072]

[0073] Where: Y i is the measured value at the i-th moment; is the predicted value at the i-th moment; is the average value of all measured values; n is the total number of measured values.

[0074] The calculation formula of accuracy is:

[0075]

[0076] The calculation formula of accuracy is:

[0077]

[0078] The calculation formula of recall rate is:

[0079]

[0080] Where: TP is the number of samples correctly predicted as positive; TN is the number of samples correctly predicted as negative; FP is the number of samples incorrectly predicted as negative as positive; FN is the number of samples incorrectly predicted as positive as negative; the positive category refers to salty tide, and the negative category refers to non-salty tide.

[0081] The calculation formula for the ratio of the predicted and measured saltwater tide times is:

[0082]

[0083] The evaluation indicators of the calculation model are as follows: NSE ranges from (-∞,1], and the closer the value is to 1, the better the fit between the model prediction and the measured data. Generally speaking, when NSE is greater than 0.6, the model fitting effect is considered to be good, so 0.6 can be used as the NSE qualified threshold; the estuary salinity threshold for judging saltwater tide is 250, that is, a saltwater tide is judged to be a saltwater tide when the salinity is greater than 250 (the calculation of the three ratios and the ratios are all calculated based on the number of saltwater tides). The closer the precision, accuracy, recall rate and the ratio of the number of saltwater tides to the measured value are to 1, the higher the accuracy of the saltwater tide prediction; according to experience, the qualified thresholds of precision, accuracy and recall rate are generally 0.7, and the qualified threshold interval of the ratio of the number of saltwater tides to the measured value is [0.8,1.2]. When all evaluation indicators reach the set qualified threshold, it means that the model test has passed.

[0084] Step 5: Salt tide prediction: Use the tested model to predict the salinity of the estuary in the target period of the study area to perform the corresponding salt tide prediction work.

[0085] Embodiment 1

[0086] This embodiment is an application example of the above method. The research area selected in this embodiment is the Modaomen Waterway, which is located in the south-central part of Guangdong Province and is one of the estuaries of the Pearl River.

[0087] This embodiment discloses a saltwater tide prediction method based on an LSTM-GRU-CNN integrated model, the method comprising the following steps:

[0088] Step 1: Construct a comprehensive data set: Obtain the measured estuary salinity data of Guangchang Pumping Station, Pinggang Pumping Station, and Zhuzhoutou Pumping Station in the study area, the flow data of Shijiao and Gaoyao, and the tide level data of Denglongshan. Select the data from October 1, 2020 to June 27, 2022 as the research data to construct a comprehensive data set.

[0089] Step 2: Data preprocessing: Preprocess the data in the comprehensive data set, including data merging, missing value filling, data normalization and format conversion; use preceding filling interpolation to complete the missing salinity, flow and tide data; then use data normalization to scale the data eigenvalues ​​to the interval [0,1]; after converting the data into a supervised learning problem, select the data from 3:00 on October 1, 2020 to 2:00 on December 1, 2021 as the training period data, and select the data from 3:00 on December 1, 2021 to 20:00 on June 27, 2022 as the test period data.

[0090] Step 3: Construct and train the LSTM-GRU-CNN integrated model: Construct the LSTM-GRU-CNN integrated model and use the training period data to train the model; use the comprehensive data of the training period with a lag time step of 18 hours as input, and the estuary salinity data of the three pumping stations with a prediction time step of 6 hours in the training period as output to train the LSTM-GRU-CNN integrated model (that is, use the hydrological data of the past 18 hours to predict the salinity data of the next 6 hours); until the model training is completed.

[0091] Step 4: Model testing and evaluation: The comprehensive data of the test period is input into the trained model to obtain the prediction results, and the prediction results are denormalized to obtain the predicted estuary salinity data, where the lag time steps and the prediction time steps are consistent with those during training; the predicted estuary salinity data are compared with the measured estuary salinity data. For the three pumping stations, the precision, accuracy and recall of the model are all greater than 0.8, and the difference between the predicted and measured ratios of the number of saltwater tides and 1 is less than or equal to 0.1, indicating that the model has high accuracy and reliability in identifying saltwater tide events; the Nash efficiency coefficient NSE is 0.91, indicating that the trained integrated model has a very good fitting effect.

[0092] Step 5: Salt tide prediction: Use the tested model to predict the salinity of the estuary in the target period of the study area to perform the corresponding salt tide prediction work.

[0093] The two deep learning models, LSTM and GRU, can better capture the long-term dependencies in time series data. The CNN model can capture local dependencies and features in time and learn nonlinear relationships through convolution operations. Therefore, the LSTM-GRU-CNN integrated model can perform time series prediction well.

[0094] In addition, this embodiment also performs two other sets of predictions: one with a lag time step of 36 hours and a prediction time step of 12 hours, and one with a lag time step of 72 hours and a prediction time step of 24 hours. The Nash efficiency coefficients NSE calculated from these two sets of prediction results are 0.91 and 0.90 respectively, and the precision, accuracy and recall of the model are all greater than 0.8, indicating that the fitting effect of the integrated model is relatively stable.

[0095] According to the comprehensive analysis of various evaluation indicators, it can be concluded that the method described in this embodiment has the characteristics of high efficiency and high precision, and can well predict the estuary salinity in the future period based on the hydrological data of the past period, thereby realizing accurate prediction of the salt tide situation in the estuary.

[0096] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, a person skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A saltwater tide prediction method based on LSTM-GRU-CNN integrated model, characterized in that: The method comprises the following steps: Step 1: Construct a comprehensive data set: Obtain the measured estuary salinity data, upstream flow data, and estuary tide data of multiple pumping stations in a certain period of time in the study area to construct a comprehensive data set; Step 2: Data preprocessing: preprocess the data in the comprehensive data set, and divide the preprocessed data into training period data and test period data in proportion; Step 3, constructing and training the LSTM-GRU-CNN integrated model: constructing the LSTM-GRU-CNN integrated model, using the training period data to train the model, setting the lag time steps and the prediction time steps, using the comprehensive data of the training period with the set lag time steps as the model input, using the measured estuary salinity data of each pumping station during the training period with the set prediction time steps as the model output, and training the model until the model training is completed; The specific process of constructing the LSTM-GRU-CNN integrated model is as follows: Step 31: Use the LSTM, Conv1D, MaxPooling1D, Flatten, GRU, Dense, and Dropout layers of the layers module in the Keras library in Python to build an LSTM model, a GRU model, and a CNN model, respectively, and set the initial parameters of the three models respectively; Step 32: Build an input layer to receive data, and connect the output of the input layer to the input of the LSTM model, GRU model, and CNN model respectively; Step 33: Use the Concatenate layer of the layers module in the Keras library to connect the outputs of the three models together and fuse the features they extracted; Step 34: Add a Dense layer to the concatenated output, whose output size is the number of features multiplied by the number of output time steps, and the activation function uses ReLU to obtain the final stacked output; Step 35: Use the Model class of the models module in the Keras library to build an LSTM-GRU-CNN integrated model. The input of the integrated model is the input layer built in step 32, and the output of the model is the stacked output obtained in step 34. Step 4, model testing and evaluation: the comprehensive data of the test period data is input into the trained integrated model, where the lag time steps and the prediction time steps are consistent with those during training, and then the model prediction results are obtained. The model prediction results are denormalized to obtain the predicted estuary salinity data, and the predicted estuary salinity data is compared with the measured estuary salinity data of each pumping station, and the evaluation index is used to determine whether the model has passed the test; Step 5: Salt tide prediction: Use the tested model to predict the salinity of the estuary in the target period of the study area to perform the corresponding salt tide prediction work.

2. A saltwater tide prediction method based on LSTM-GRU-CNN integrated model according to claim 1, characterized in that: The preprocessing process described in step 2 includes data merging, missing value filling, data normalization and format conversion; The data is merged to generate a time series containing hourly time points according to the minimum and maximum values ​​of the series, and the original data and the time series are merged using left connection to ensure that all time points are included; The missing value filling is to use the preceding filling method to complete the missing salinity, flow and tide level data, that is, to fill the missing value with the nearest non-missing value before the missing value, so as to obtain complete data; The data normalization is to use the MinMaxScaler class in the scikit-learn library in Python to perform data normalization and scale the data feature values ​​to the interval [0,1]; The format conversion converts the original data into a format for supervised learning, that is, the data is divided into two parts: input features and output labels, and the corresponding relationship between them is maintained.

3. A saltwater tide prediction method based on LSTM-GRU-CNN integrated model according to claim 1, characterized in that: The ratio of the training data to the test data in step 2 is 7:

3.

4. A saltwater tide prediction method based on LSTM-GRU-CNN integrated model according to claim 1, characterized in that: The specific process of training the model in step 3 until the model training is completed is: (1) First, a common loss function is used. According to the prediction results of the initial model using the common loss function, weights are set for the multiple pumping stations described in step 1 to adjust the model's emphasis on different pumping stations; (2) Define a weighted loss function and use the mean square logarithmic error to calculate the relative error between the predicted value and the true value of each pumping station; since the data of multiple pumping stations are included in the comprehensive data set, the data of the corresponding row and column positions of each pumping station are extracted through Python slicing operations; (3) Calculate the loss of the corresponding pumping station according to the weight, add up the weighted sum of the losses of each pumping station, and get the total loss value as the return value of the function. After the training is completed, output the change curve of the weighted loss function and analyze the image. When the function converges, it means that the model training is completed.

5. A saltwater tide prediction method based on LSTM-GRU-CNN integrated model according to claim 1, characterized in that: The specific process of using evaluation indicators to determine whether the model has passed the test in step 4 is as follows: The evaluation indicators include: Nash efficiency coefficient NSE, precision, accuracy, recall rate and the ratio of the predicted and measured saltwater tide times of each pumping station; The calculation formula of NSE is: Where: Y i is the measured value at the i-th moment; is the predicted value at the i-th moment; is the average value of all measured values; n is the total number of measured values; The calculation formula of accuracy is: The calculation formula of accuracy is: The calculation formula of recall rate is: Where: TP is the number of samples correctly predicted as positive; TN is the number of samples correctly predicted as negative; FP is the number of samples incorrectly predicted as negative as positive; FN is the number of samples incorrectly predicted as positive as negative; positive refers to salty tide, negative refers to non-salty tide; The calculation formula for the ratio of the predicted and measured saltwater tide times is: Calculate each evaluation index of the model and determine whether each evaluation index reaches the qualified threshold. When each evaluation index reaches the set qualified threshold, it means that the model test has passed.

6. A saltwater tide prediction method based on LSTM-GRU-CNN integrated model according to claim 5, characterized in that: The NSE qualified threshold is 0.6; the qualified thresholds of the precision, accuracy and recall are all 0.7; the qualified threshold interval of the ratio of the predicted and measured saltwater tide times is [0.8, 1.2].

Images