Dam body water seepage amount prediction method based on multi-scale convolutional neural network and bidirectional long-short-term memory neural network

Abstract

The invention discloses a dam body water seepage amount prediction method based on a multi-scale convolutional neural network and a bidirectional long-short-term memory neural network, and the method comprises the steps: 1, carrying out the data collection and preprocessing: 2, building an MS-CNN neural network, and extracting the features of spatial data in monitoring data; step 3, establishing a Bi-LSTM neural network, and extracting features of time data in the monitoring data; 4, model fusion and training: taking the output of the MS-CNN neural network and the time sequence data as the input of a Bi-LSTM neural network, and realizing the lap joint of the two neural networks to obtain a complete water seepage prediction model; training the model through processed dam body water seepage monitoring data, optimizing parameters of the model through a back propagation algorithm, and finely adjusting hyper-parameters to improve the recognition capability of the model; the method improves the accuracy and robustness of the whole prediction model, achieves the high-precision prediction of the water seepage amount of the dam body and the early warning of the abnormal water seepage amount, and provides powerful support for the safety monitoring and maintenance of the dam.

Description

Technical Field

[0001] The invention relates to the technical field of dam body seepage prediction, in particular to a dam body seepage prediction method based on multi-scale convolutional neural network and bidirectional long short-term memory neural network. Background Art

[0002] As an important water conservancy project facility, the safety and stability of the dam are of vital importance. However, during the long-term operation of the dam, affected by multiple factors such as the natural environment, material aging, design and construction, water seepage in the structural joints and water seepage in the bedrock cracks of the dam body often occurs. These problems not only affect the normal operation of the dam, but may also cause serious safety accidents. Therefore, it is very necessary to monitor and predict its seepage.

[0003] At present, the prediction of dam seepage mainly relies on traditional physical models and empirical formulas, but these methods often have problems such as low prediction accuracy and poor adaptability. Especially for complex and changeable seepage conditions, traditional methods are difficult to accurately capture the dynamic characteristics of seepage in space and time, and it is difficult to make truly effective predictions. Specifically, the current traditional methods have the following technical problems: (1) Difficulty in feature extraction: In the process of dam seepage prediction, monitoring data usually contains complex spatial distribution information and time series information, accompanied by the influence of noise and outliers. These factors make it difficult for traditional methods to accurately and effectively extract features that have a decisive impact on seepage prediction from raw data.

[0004] (2) Difficulty in combining spatial and temporal characteristics: The change in dam seepage is not only affected by the current spatial distribution characteristics, but also closely related to the time series characteristics of the past and future. However, traditional methods can only process simple time data and it is difficult to effectively combine these two characteristics for comprehensive analysis.

[0005] (3) Low prediction accuracy of dam seepage: The accuracy of traditional methods and empirical formulas in predicting dam seepage is often difficult to meet actual needs, and the prediction accuracy is low. Summary of the invention

[0006] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a dam seepage prediction method based on multi-scale convolutional neural network and bidirectional long short-term memory neural network to solve the problems raised in the background technology.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for predicting dam seepage based on multi-scale convolutional neural network and bidirectional long short-term memory neural network, comprising the following steps: Step 1: Data collection and preprocessing: Step 2: Build the MS-CNN neural network to extract the features of spatial data in the monitoring data; Step 3: Build a Bi-LSTM neural network to extract the features of time data in the monitoring data; Step 4: Model fusion and training: The output of the MS-CNN neural network and the time series data are used as the input of the Bi-LSTM neural network to realize the overlap of the two neural networks and obtain a complete seepage prediction model; it is trained with the processed dam seepage monitoring data, the model parameters are optimized through the back propagation algorithm, and the hyperparameters are fine-tuned to improve the model recognition ability, so that the model can finally achieve the dam seepage prediction task.

[0008] Preferably, the step 1 specifically includes the following process: The dam monitoring data is collected, including time series data of seepage, water level, temperature, and humidity, as well as spatial data of seepage crack location, elevation, length, and width. The collected data is then cleaned to remove outliers and missing values, and interpolation or estimation is performed. Finally, the time series data is processed by wavelet transformation to capture characteristic information at different time scales. At the same time, all the data is normalized so that it can be input into the model for training and prediction.

[0009] Preferably, the specific process of building the MS-CNN neural network in step 2 is as follows: Build an MS-CNN neural network containing multiple scale convolutional layers, which includes an input layer, multiple multi-scale convolutional layers and a feature fusion layer; each scale convolutional layer can extract feature information of different spatial dimensions, and the feature fusion layer fuses the feature information extracted by different scale convolutional layers, so that the network can simultaneously utilize features of different scales.

[0010] Preferably, the convolution calculation formula is:

[0011] Among them: The feature map of the i-th layer, is the convolution kernel of the i-th layer, is the new calculation result, Represents convolution calculation.

[0012] Preferably, the specific process of building the Bi-LSTM neural network in step 3 is as follows: A Bi-LSTM neural network is constructed, which includes an input layer, a bidirectional LSTM layer, a fully connected layer, and an output layer; the bidirectional LSTM layer consists of a forward LSTM and a reverse LSTM, and these two LSTM networks extract the forward and reverse features of the time series respectively.

[0013] Preferably, the output of the bidirectional LSTM layer is the output of the combined forward and reverse LSTMs, and the calculation formula is:

[0014] in: is the output of the bidirectional LSTM layer, is the output result of the forward LSTM, is the output of the reverse LSTM.

[0015] Preferably, the step 4 specifically includes the following process: S4.1: The MS-CNN neural network and the Bi-LSTM neural network are overlapped to form a complete water seepage prediction model. During the overlap process, the output result of the MS-CNN neural network is dimensionally adjusted through the output layer to make it consistent with the time dimension of the time series data set. The output result is combined with the time series data set to form new input data, which is input into the Bi-LSTM neural network, and finally the prediction result of water seepage is output. By overlapping the two neural network models, a prediction model for dam water seepage is formed. S4.2: The processed spatial data set and time series data set are divided into training set, validation set and test set in a ratio of 6:2:2. The training set and validation set are used to train the parameters of the model and adjust the hyperparameters based on them. The test set is used to test the final effect of the model in predicting water seepage. S4.3: Train the network parameters of the MS-CNN neural network and the Bi-LSTM neural network, and manually adjust their network structure and hyperparameters to achieve accurate prediction, analysis and early warning of water seepage in the dam structure joints and bedrock cracks.

[0016] Beneficial effects of the present invention: (1) Enhanced model robustness: The multi-scale feature extraction mechanism of the MS-CNN neural network enables the model to automatically adapt to spatial features of different dimensions and scales, while the Bi-LSTM neural network can effectively handle long-term dependencies and uncertainties in time series. The present invention combines the two to not only improve the model's adaptability to complex data structures, but also enhance the model's robustness to noise and outliers, ensuring stable prediction performance under various environmental conditions.

[0017] (2) Significantly improve prediction accuracy: By integrating the spatial feature extraction capability of MS-CNN and the time series feature analysis capability of Bi-LSTM, the present invention can comprehensively capture the changing impact of the spatial distribution of dam seepage and its dynamic trend over time, thereby significantly improving the accuracy of the prediction model. Its accurate prediction of seepage is crucial for dam safety monitoring and subsequent maintenance.

[0018] (3) Realizing high-precision early warning function: Through high-precision prediction, the present invention can timely detect and warn of the occurrence of abnormal water seepage, provide data support for the safety management and maintenance of the dam, and provide dam managers with sufficient response time; this helps to detect problems in a timely manner and take measures to ensure the structural safety and stable operation of the dam.

[0019] (4) Improving the intelligence of dam monitoring: The present invention can be integrated into the existing monitoring system to realize automatic monitoring, analysis and early warning of water seepage in the dam structure joints and other areas, thereby improving the intelligence of dam management and improving work efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a flow chart of the present invention; Figure 2 It is a schematic diagram of the algorithm structure of the present invention; Figure 3 It is a schematic diagram of the specific structure of the Bi-LSTM network layer in the Bi-LSTM neural network of the present invention. DETAILED DESCRIPTION

[0021] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0022] like Figure 1-3 As shown, a dam seepage prediction method based on multi-scale convolutional neural network and bidirectional long short-term memory neural network includes the following steps: Step 1: Data collection and preprocessing: S1.1: Collect water seepage monitoring data of dam structure joints or bedrock cracks, mainly including five time series data: water seepage volume, water level, precipitation, temperature, and humidity, and four spatial data: water seepage location (expressed by pile number), elevation, crack length, and crack width.

[0023] S1.2: Process the spatial data and time series data in a unified dimension so that they can be used as input data for the water seepage prediction model. Because the dimension of spatial data is generally unchanged in the time series and has only one value, while the time series data has a lot of data in the time dimension, the two dimensions are inconsistent. Therefore, they need to be processed so that the data format of each monitoring point is in the form of time as the vertical coordinate, and spatial data and time series data as the horizontal coordinate, and there are corresponding spatial data and time data at each time point. Since space and time are basically unchanged at a monitoring point, this can be achieved by copying the spatial data.

[0024] S1.3: Remove outliers and denoise the raw data of the unified time dimension, such as the amount of water seepage that may be recorded incorrectly by humans. Then use the Random Forest algorithm to interpolate the gaps in the raw data. At the same time, perform wavelet transformation on the time series data to capture the characteristic information of different time scales. Finally, divide the spatial data and time series data into two data sets, normalize them separately, and use them as training data sets for subsequent models.

[0025] Step 2: Build the MS-CNN neural network to extract the features of spatial data in the monitoring data; The MS-CNN neural network is constructed, which mainly includes: 1 input layer, 3 multi-scale convolution layers and 1 feature fusion layer. Its core calculation is the multi-scale convolution layer. Convolution layers of different depths, each layer uses convolution kernels of different sizes to capture spatial features of different scales. Its main convolution calculation formula is:

[0026] Among them: The feature map of the i-th layer, is the convolution kernel of the i-th layer, is the new calculation result, Represents convolution calculation.

[0027] Step 3: Build a Bi-LSTM neural network to extract the features of time data in the monitoring data; The Bi-LSTM neural network is constructed, which mainly includes: 1 input layer, 6 bidirectional LSTM layers, 3 fully connected layers and 1 output layer. The core is the bidirectional LSTM layer, which consists of a forward LSTM and a reverse LSTM layer. The output of the bidirectional LSTM layer is the output of the forward and reverse LSTM layers. The calculation formula is:

[0028] in: is the output of the bidirectional LSTM layer, is the output result of the forward LSTM, is the output of the reverse LSTM.

[0029] Step 4: Model fusion and training: The output of the MS-CNN neural network and the time series data are used as the input of the Bi-LSTM neural network to realize the overlap of the two neural networks and obtain a complete seepage prediction model; it is trained with the processed dam seepage monitoring data, the model parameters are optimized through the back propagation algorithm, and the hyperparameters are fine-tuned to improve the model recognition ability, so that the model can finally achieve the dam seepage prediction task.

[0030] S4.1: Overlap the MS-CNN neural network and the Bi-LSTM neural network to form a complete water seepage prediction model. During the overlapping process, the output result of the MS-CNN neural network is dimensionally adjusted through the output layer to make it consistent with the time dimension of the time series data set. The output result is combined with the time series data set to form new input data, which is input into the Bi-LSTM neural network, and finally the prediction result of water seepage is output. By overlapping the two neural network models, a prediction model for dam water seepage is formed.

[0031] S4.2: The processed spatial dataset and time series dataset are divided into training set, validation set and test set in the ratio of 6:2:2. The training set and validation set are used to train the parameters of the model and adjust the hyperparameters based on them. The test set is used to test the final effect of the model in predicting the seepage volume. When splitting the data, it is necessary to pay attention to the strict comparison between the spatial dataset and the time series dataset, and they cannot be split arbitrarily.

[0032] S4.3: The structure of the dam seepage prediction model is as follows: Figure 2 As shown in the figure, the network parameters of the MS-CNN neural network and the Bi-LSTM neural network are mainly trained, and their network structure and hyperparameters are manually adjusted to achieve the best prediction effect. The final model can accurately predict the seepage volume of the dam structure joints and bedrock cracks, and realize analysis and early warning, providing data support for the safety management and maintenance of the dam.

[0033] In the above technical solution, Convolutional neural network: Convolutional neural network has powerful feature extraction capabilities and has a strong ability in water level prediction. By extracting spatial features related to water level, the prediction of future water level can be achieved. Although these methods are mainly aimed at water level rather than seepage, their successful application in spatial feature extraction provides an important technical basis and inspiration for the MS-CNN part of this invention.

[0034] BP-RBF neural network: This network model is composed of BP neural network and radial basis function (RBF) neural network, and is widely used in the field of seepage prediction. This model combines the strong prediction ability of BP neural network and the fast convergence characteristics of RBF neural network, improves the accuracy and efficiency of seepage prediction, and provides ideas and references for model construction.

[0035] MS-CNN neural network: Multi-scale convolutional neural network is a specially designed convolutional neural network that aims to capture different levels of feature information in the input data through convolution kernels of different scales. In the scenario of dam seepage prediction, MSCNN can extract the multi-scale features of the dam and its surrounding environment in space, and improve the accuracy of deep water prediction.

[0036] Bi-LSTM neural network: Bidirectional long short-term memory neural network is a special recurrent neural network that can simultaneously capture the forward and backward dependencies in time series data by combining the information flow in the two directions of forward LSTM and backward LSTM. This structure makes Bi-LSTM very good at processing time series prediction. In the scenario of dam seepage prediction, the seepage is not only affected by various factors at the current moment, but also closely related to past and future factors. For example, factors such as rainfall and water level may affect the seepage, and these factors often have time series correlation. Therefore, Bi-LSTM has great advantages in predicting the seepage of dam structure joints.

[0037] A detailed description of the specific training process of the model is given: (1) 60 seepage points are selected, and 300 time points are selected for each point. The size of the spatial data set is 4×18000, and the size of the time series data set is 5×18000. They are divided into training sets of 4×10800 and 5×10800, validation sets of 4×3600 and 5×3600, and test sets of 4×3600 and 5×3600, respectively, in a ratio of 6:2:2.

[0038] (2) The spatial data training set is used as input data with a size of 4×10800, and is calculated in each layer of the MS-CNN neural network in turn. First, it enters the input layer, whose size is 4×10800, and then enters three multi-scale convolutional layers. The convolution kernel size of the first multi-scale convolutional layer is set to 3×1, the number of convolution kernels is 8, and the padding is 1. After passing, the output data becomes 8×10800; the convolution kernel size of the second multi-scale convolutional layer is set to 6×1, the number of convolution kernels is 16, and the padding is 3. After passing, the output data becomes 16×10800; the convolution kernel size of the third multi-scale convolutional layer is set to 9×1, the number of convolution kernels is 32, and the padding is 4. After passing, the output data becomes 32×10800. Then it enters the feature fusion layer, and the output result becomes 1×10800. This output data is called the spatial feature output set. After the three multi-scale convolutional layers, add the Relu6 activation function, and add the PRelu activation function in the feature fusion layer.

[0039] (3) The spatial feature dataset and time series dataset output by the MS-CNN neural network are combined into a new dataset with a size of 6×10800. This new dataset is used as the input of the Bi-LSTM neural network. Since the input format of the Bi-LSTM network layer is relatively special, the input dataset needs to be configured and its input shape is defined as (6, 1). The number of units in the five Bi-LSTM network layers is defined as 32, 64, 128, 64, and 32, respectively, and the final output data size is 32×10800. After that, it passes through three fully connected layers with sizes of 16×10800, 8×10800, and 4×10800, respectively. After the three fully connected layers, the Relu6 activation function is connected respectively. In the last output layer, the activation function is the PRelu activation function, and the output result is 1×10800, which is the predicted water seepage result of the model.

[0040] The above is the forward propagation process of model training. Through forward propagation, the predicted water seepage data can be obtained. At this time, the predicted data of the model is obviously impossible to be close to the actual water seepage data. Therefore, it is necessary to train the parameters of the water seepage prediction model through the back propagation process and adjust the hyperparameters of the model.

[0041] (4) In the back propagation process, Huber Loss is selected as the loss function of the present invention. The Huber Loss function combines the advantages of the mean square error loss function (MSE) and the mean absolute error loss function (MAE), which helps the model to fit the data more accurately, resist the influence of outliers, and improve the prediction accuracy. Its main formula is:

[0042] in, is the true value, is the predicted value of the model, is the threshold parameter of Huber Loss, when When , it becomes MSE; when When , it becomes similar to MAE, so Huber Loss has the advantages of both MSE and MAE.

[0043] Adam is selected as the optimizer. It is an optimization algorithm with adaptive learning rate, which has the characteristics of adaptability, high efficiency and stability. Adam optimizer uses exponential decay method to update the learning rate. Its main formula is:

[0044] Is a positive number less than 1, used to control the speed of learning rate decay. is the current iteration number, is the decay step size, is the decayed learning rate, is the initial learning rate, is the decay coefficient, and [] is rounded down.

[0045] In back propagation, the model after forward propagation is used to verify the validation set, that is, the predicted output is obtained using the input of the validation set. The predicted output is compared with the actual seepage value. The comparison between the two is the operation of the loss function. The comparison result is used to adjust the parameters of the model through the optimizer, etc., so as to carry out training.

[0046] (5) Forward propagation and backward propagation constitute a complete training. By repeating forward propagation and backward propagation many times, the weight parameters of the seepage prediction model are continuously adjusted until the best prediction accuracy is achieved. At the same time, the early stopping mechanism is used to obtain the model with the highest prediction accuracy, which is the final dam seepage prediction model.

[0047] Finally, the test set is used in the trained dam seepage prediction model to obtain the corresponding output, thereby completing the entire content and task of the present invention, that is, realizing the accurate prediction of seepage in areas such as bedrock cracks in the dam structure. Finally, an analysis and early warning algorithm is added to the model to realize the analysis of dam seepage, set the seepage threshold, and automatically initiate an alarm when it exceeds the set value. This provides data support and technical support for the intelligent forecasting and early warning of dam seepage, subsequent processing, and other work.

[0048] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A dam seepage prediction method based on multi-scale convolutional neural network and bidirectional long short-term memory neural network, Features: The following steps are involved: Step 1: Data collection and preprocessing: Step 2: Build the MS-CNN neural network to extract the features of spatial data in the monitoring data; Step 3: Build a Bi-LSTM neural network to extract the features of time data in the monitoring data; Step 4: Model fusion and training: The output of the MS-CNN neural network and the time series data are used as the input of the Bi-LSTM neural network to realize the overlap of the two neural networks and obtain a complete seepage prediction model; it is trained with the processed dam seepage monitoring data, the model parameters are optimized through the back propagation algorithm, and the hyperparameters are fine-tuned to improve the model recognition ability, so that the model can finally achieve the dam seepage prediction task.

2. According to claim 1, a dam seepage prediction method based on multi-scale convolutional neural network and bidirectional long short-term memory neural network, Features: The step 1 specifically includes the following process: The dam monitoring data is collected, including time series data of seepage, water level, temperature, and humidity, as well as spatial data of seepage crack location, elevation, length, and width. The collected data is then cleaned to remove outliers and missing values, and interpolation or estimation is performed. Finally, the time series data is processed by wavelet transformation to capture characteristic information at different time scales. At the same time, all the data is normalized so that it can be input into the model for training and prediction.

3. According to claim 1, a dam seepage prediction method based on multi-scale convolutional neural network and bidirectional long short-term memory neural network, Features: The specific process of building the MS-CNN neural network in step 2 is as follows: Build an MS-CNN neural network containing multiple scale convolutional layers, which includes an input layer, multiple multi-scale convolutional layers and a feature fusion layer; each scale convolutional layer can extract feature information of different spatial dimensions, and the feature fusion layer fuses the feature information extracted by different scale convolutional layers, so that the network can simultaneously utilize features of different scales.

4. According to claim 3, a dam seepage prediction method based on multi-scale convolutional neural network and bidirectional long short-term memory neural network, Features: The convolution calculation formula is: Among them: The feature map of the i-th layer, is the convolution kernel of the i-th layer, is the new calculation result, Represents convolution calculation.

5. According to claim 1, a dam body seepage prediction method based on multi-scale convolutional neural network and bidirectional long short-term memory neural network, Features: The specific process of building the Bi-LSTM neural network in step 3 is as follows: A Bi-LSTM neural network is constructed, which includes an input layer, a bidirectional LSTM layer, a fully connected layer, and an output layer; the bidirectional LSTM layer consists of a forward LSTM and a reverse LSTM, and these two LSTM networks extract the forward and reverse features of the time series respectively.

6. A dam seepage prediction method based on multi-scale convolutional neural network and bidirectional long short-term memory neural network according to claim 5, Features: The output of the bidirectional LSTM layer is the combined output of the forward and reverse LSTMs, and its calculation formula is: in: is the output of the bidirectional LSTM layer, is the output result of the forward LSTM, is the output of the reverse LSTM.

7. The method for predicting dam seepage based on multi-scale convolutional neural network and bidirectional long short-term memory neural network according to claim 1, Features: The step 4 specifically includes the following process: S4.1: The MS-CNN neural network and the Bi-LSTM neural network are overlapped to form a complete water seepage prediction model. During the overlap process, the output result of the MS-CNN neural network is dimensionally adjusted through the output layer to make it consistent with the time dimension of the time series data set. The output result is combined with the time series data set to form new input data, which is input into the Bi-LSTM neural network, and finally the prediction result of water seepage is output. By overlapping the two neural network models, a prediction model for dam water seepage is formed. S4.2: The processed spatial data set and time series data set are divided into training set, validation set and test set in a ratio of 6:2:

2. The training set and validation set are used to train the parameters of the model and adjust the hyperparameters based on them. The test set is used to test the final effect of the model in predicting water seepage. S4.3: Train the network parameters of the MS-CNN neural network and the Bi-LSTM neural network, and manually adjust their network structure and hyperparameters to achieve accurate prediction, analysis and early warning of water seepage in the dam structure joints and bedrock cracks.

Images