A Neural Network Reservoir Water Level Prediction Method Based on Ensemble Learning

Abstract

The present invention discloses a neural network reservoir water level prediction method based on ensemble learning. The method comprises the following steps: selecting historical water level data from long-term inflows at spatially adjacent water level stations to form a data set; performing numerical transformation (i.e., data normalization) on the data set to obtain a standard data set, and dividing the standard data set into a training set, a validation set, and a test set; constructing a neural network model based on ensemble learning and defining a loss function; selecting data from the standard data set to train the neural network model, and performing parameter verification and parameter adjustment on the trained network model; employing an Epoch-Wise Bagging ensemble learning strategy to select M optimal network models after verification for testing, and finally predicting the spatial distribution of the reservoir's real-time water level within the next day based on the real-time water inflow of each water level station. This method not only considers the temporal correlation of reservoir water level inflow data, but also the spatial correlation between each water level station, thereby better exploring the law of reservoir water level variation.

Description

Technical Field

[0001] The present invention belongs to the technical field of real-time reservoir water level prediction, and in particular relates to a neural network reservoir water level prediction method based on ensemble learning. Background Art

[0002] Reservoir water levels are fundamental to maintaining the structure, function, and integrity of their ecosystems. The normal water level determines the reservoir's size, efficiency, and regulation, as well as the size and type of hydraulic structures and reservoir flooding losses. It is the most crucial characteristic of reservoir operation. Furthermore, as the most direct and important indicator of reservoir water regime changes, water level holds significant practical significance in modern lake research. Developing effective reservoir water level simulation methods has become a crucial and urgent task facing current research.

[0003] A commonly used reservoir water level prediction method uses historical water level change data from a reservoir water level monitoring system as input, extracting its statistical characteristics (such as mean, variance, maximum, and minimum values) to predict water level trends over a specific time period. Current prediction results show low accuracy and poor stability due to the influence of inflow from adjacent water level stations within the reservoir. In recent years, with the widespread adoption of big data technologies and machine learning algorithms, the application of artificial intelligence methods to real-time reservoir water level prediction has become a research hotspot.

[0004] Historical reservoir water level data is time series data. Water flowing into adjacent water level stations is regulated by the reservoir, resulting in inconsistent changes in water level and flow, and a significant phase lag. Artificial intelligence methods, such as machine learning and deep learning, can effectively explore the potential temporal relationship between water flow and reservoir level at adjacent water level stations, enabling high-precision, real-time prediction of reservoir water levels at a specific moment in the future. Summary of the Invention

[0005] Purpose of the invention: Based on the technical problems raised above, a neural network reservoir water level prediction method based on ensemble learning is provided. The present invention constructs a neural network reservoir water level prediction model based on convolutional neural networks and long-short-term memory networks based on the historical water level data collected from adjacent water level stations in the reservoir space, and uses the Epoch-WiseBagging ensemble learning strategy to optimize the model accuracy. On the one hand, a two-dimensional convolutional neural network is used to capture a certain correlation between water level values between water level stations at different locations in space, thereby improving the information extraction ability of the model; on the other hand, a long-short-term memory network is used to learn the time dependence and change trend of the reservoir water level, thereby improving the accuracy of the model water level prediction.

[0006] Technical solution: In order to solve the above technical problems, the present invention proposes a neural network reservoir water level prediction method based on ensemble learning, which includes the following steps:

[0007] Step 1: Collect historical water level data of each adjacent water level station in the reservoir space within a preset time period to form reservoir water level sample data;

[0008] Step 2: Perform numerical transformation on the reservoir water level sample data in step 1, i.e., data normalization, to obtain a standard data set, and divide the standard data set into a training set, a validation set, and a test set;

[0009] Step 3: Build a hybrid model based on convolutional neural network and long short-term memory network, and define the model loss function;

[0010] Step 4: Use the training set in step 2 to train the neural network model built in step 3, and use the gradient descent strategy to optimize the loss function of step 3;

[0011] Step 5: Use the validation set in step 2 to adjust the parameters of the neural network model, determine the network weight parameters of the model through multiple validations, and save the M optimal models with the highest prediction accuracy during the validation process;

[0012] Step 6: According to the Epoch-Wise Bagging ensemble learning strategy, select the M best network models verified by the validation set in step 5, and test these M best network models using the test set in step 2;

[0013] Step 7: Input the real-time water inflow of each water level station into M optimal network models, and calculate the M output results corresponding to these M optimal network models using the weighted average method to obtain the final model output to predict the real-time water level spatial distribution of the reservoir within the next 1 day.

[0014] Furthermore, in step 1, the water level historical data are collected and formed into an input sample data set for predicting the reservoir water level, and its shape is expressed as [N, t, k], where N represents the total sample size of the input data of the reservoir water level, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station for t consecutive days.

[0015] Furthermore, the data in step 2 is normalized as follows:

[0016]

[0017] Among them, X cur are the values of t and k that need to be normalized; X min are the minimum values of t and k respectively; Xmax are the maximum values of t and k respectively;

[0018] The normalized input data samples are divided into training set, validation set and test set in the ratio of 8:1:1.

[0019] Furthermore, in step 3, the convolutional neural network and long short-term memory hybrid model specifically includes the following layers, in order: an input layer, three two-dimensional convolutional layers, a first LSTM layer, a second LSTM layer, and an output layer;

[0020] The data of the input layer is [N, t, k], where N represents the sample size of the reservoir water level data, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station within t consecutive days; there are three two-dimensional convolution layers, denoted as C1, C2, and C3, respectively, where the convolution layer is expressed as: C = f(wx+b), where f is the activation function, and all three layers are ReLU activation functions. In order to match the two-dimensional convolution operation, the input layer is expanded by one dimension, and the expanded input shape is [N, t, k, 1]. By setting the size of the convolution kernel and the padding operation, the output after the three-layer two-dimensional convolution layer is further obtained, and its output dimension is [N, m, k, 256], where m represents the value after the three-layer convolution, which includes the time change law of the water level within t days; the number of output channels is 256, which integrates the water level change data between k water level stations;

[0021] The output dimensions after the three convolution layers are combined to obtain an output shape of [N, m*k*256] to match the input of the first LSTM layer. The output shape after the first LSTM layer is [N, 256], where the number of neurons in the middle hidden layer is 256. Finally, the output is sent to the second LSTM layer and the output shape is [N, 512], where the number of output neurons is 512.

[0022] The output layer is a fully connected layer that receives the output from the second LSTM layer as its input and obtains an output of shape [N, 5], where 5 represents the daily average water level data of the five downstream water level stations in the next day;

[0023] The loss function is:

[0024]

[0025] Among them, Y obs is the actual reservoir water level, Y pre is the predicted value of the above neural network model, and N represents the total number of samples.

[0026] Furthermore, in steps 4 and 5, the Adam optimizer based on the gradient descent strategy is used to optimize the weight parameters during the network training process. The batch processing parameter batch_size is set to 200, the number of training rounds is set to 100, the learning rate is set to 0.0001, and the Dropout regularization method is used to reduce network overfitting. After the network model is built based on the training set, the network hyperparameters are adjusted on the validation set. The network weight parameters of the model are determined through multiple validations, and the M optimal models with the highest prediction accuracy during the validation process are saved.

[0027] Furthermore, in step 6, the Epoch-Wise Bagging ensemble learning strategy includes: building a convolutional neural network and a long short-term memory hybrid model based on the training set, adjusting the weight parameters of the network through the validation set and saving M optimal accuracy models as base learners, and then selecting these M optimal base learners for ensemble learning testing and outputting the prediction results of the M optimal base learners.

[0028] The inventors propose a neural network reservoir water level prediction system based on ensemble learning, which includes the following steps:

[0029] Data collection module: collects historical water level data of each adjacent water level station in the reservoir space within a preset time period to form reservoir water level sample data;

[0030] Data processing module: Perform numerical transformation on reservoir water level sample data, i.e. data normalization, to obtain a standard data set, and divide the standard data set into training set, validation set and test set;

[0031] Neural network building module: Build a hybrid model based on convolutional neural network and long short-term memory network, and define the model loss function;

[0032] Neural network training module: Use the training set to train the constructed neural network model and use the gradient descent strategy to optimize the loss function;

[0033] Neural network optimization module: Use the validation set to adjust the parameters of the neural network model, determine the network weight parameters of the model through multiple validations, and save the M optimal models with the highest prediction accuracy during the validation process;

[0034] Optimal network model selection module: Based on the Epoch-Wise Bagging ensemble learning strategy, select M optimal network models verified by the validation set, and use the test set to test these M optimal network models;

[0035] Water level distribution prediction module: The real-time water inflow of each water level station is input into M optimal network models, and the M output results corresponding to these M optimal network models are calculated by weighted averaging to obtain the final model output to predict the real-time water level spatial distribution of the reservoir within the next 1 day.

[0036] Preferably, in the data collection module, the water level historical data is collected and formed into an input sample data set for predicting the reservoir water level, and its shape is expressed as [N, t, k], where N represents the total sample size of the input data of the reservoir water level, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station for t consecutive days.

[0037] Preferably, in the data processing module, the data is normalized as follows:

[0038]

[0039] Among them, X cur are the values of t and k that need to be normalized; X min are the minimum values of t and k respectively; X max are the maximum values of t and k respectively;

[0040] The normalized input data samples are divided into training set, validation set and test set in the ratio of 8:1:1.

[0041] Preferably, in the neural network building module, the convolutional neural network and the long short-term memory hybrid model specifically include the following layers, in order: input layer, three two-dimensional convolutional layers, the first LSTM layer, the second LSTM layer, and the output layer;

[0042] The data of the input layer is [N, t, k], where N represents the sample size of the reservoir water level data, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station within t consecutive days; there are three two-dimensional convolution layers, denoted as C1, C2, and C3, respectively, where the convolution layer is expressed as: C = f(wx+b), where f is the activation function, and all three layers are ReLU activation functions. In order to match the two-dimensional convolution operation, the input layer is expanded by one dimension, and the expanded input shape is [N, t, k, 1]. By setting the size of the convolution kernel and the padding operation, the output after the three-layer two-dimensional convolution layer is further obtained, and its output dimension is [N, m, k, 256], where m represents the value after the three-layer convolution, which includes the time change law of the water level within t days; the number of output channels is 256, which integrates the water level change data between k water level stations;

[0043] The output dimensions after the three convolution layers are combined to obtain an output shape of [N, m*k*256] to match the input of the first LSTM layer. The output shape after the first LSTM layer is [N, 256], where the number of neurons in the middle hidden layer is 256. Finally, the output is sent to the second LSTM layer and the output shape is [N, 512], where the number of output neurons is 512.

[0044] The output layer is a fully connected layer that receives the output from the second LSTM layer as its input and obtains an output of shape [N, 5], where 5 represents the daily average water level data of the five downstream water level stations in the next day;

[0045] The loss function is:

[0046]

[0047] Among them, Y obs is the actual reservoir water level, Y pre is the predicted value of the above neural network model, and N represents the total number of samples.

[0048] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0049] The calculation accuracy of the model of the present invention can be continuously improved with the accumulation of reservoir operation data, and it has low dependence on the accuracy of boundary conditions. It can effectively overcome the problems of low calculation accuracy and slow network convergence of traditional hydrodynamics and machine learning methods. In the flood season scheduling of reservoirs, it can accurately predict the real-time water level in the future, providing strong technical support for flood control and benefit scheduling decisions of large reservoirs or cascade reservoirs. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of the method flow of the present invention;

[0051] Figure 2 It is a schematic structural diagram of the reservoir water level prediction model of the present invention;

[0052] Figure 3 A schematic diagram of the spatial relationship of water level stations provided in an embodiment of the present invention;

[0053] Figure 4 Schematic diagram of the training process of the reservoir water level prediction model according to an embodiment of the present invention;

[0054] Figure 5 Schematic diagram comparing the prediction results of the reservoir water level prediction model according to an embodiment of the present invention and other models. DETAILED DESCRIPTION

[0055] The technical solutions and effects of the present invention will be described in detail below with reference to the accompanying drawings and specific implementations.

[0056] The present invention proposes a neural network reservoir water level prediction method based on ensemble learning, which includes the following steps:

[0057] Step 1: Collect historical water level data of each adjacent water level station in the reservoir space within a preset time period to form reservoir water level sample data;

[0058] Step 2: Perform numerical transformation on the reservoir water level sample data in step 1, i.e., data normalization, to obtain a standard data set, and divide the standard data set into a training set, a validation set, and a test set;

[0059] Step 3: Build a hybrid model based on convolutional neural network and long short-term memory network, and define the model loss function;

[0060] Step 4: Use the training set in step 2 to train the neural network model built in step 3, and use the gradient descent strategy to optimize the loss function of step 3;

[0061] Step 5: Use the validation set in step 2 to adjust the parameters of the neural network model, determine the network weight parameters of the model through multiple validations, and save the M optimal models with the highest prediction accuracy during the validation process;

[0062] Step 6: According to the Epoch-Wise Bagging ensemble learning strategy, select the M best network models verified by the validation set in step 5, and test these M best network models using the test set in step 2;

[0063] Step 7: Input the real-time water inflow of each water level station into M optimal network models, and calculate the M output results corresponding to these M optimal network models using the weighted average method to obtain the final model output to predict the real-time water level spatial distribution of the reservoir within the next 1 day.

[0064] In step 1, the water level historical data are collected and formed into an input sample data set for predicting the reservoir water level, and its shape is expressed as [N, t, k], where N represents the total sample size of the input data of the reservoir water level, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station for t consecutive days.

[0065] The data in step 2 is normalized to:

[0066]

[0067] Among them, X cur are the values of t and k that need to be normalized; Xmin are the minimum values of t and k respectively; X max are the maximum values of t and k respectively;

[0068] The normalized input data samples are divided into training set, validation set and test set in the ratio of 8:1:1.

[0069] In step 3, the convolutional neural network and long short-term memory hybrid model specifically includes the following layers, in order: input layer, three two-dimensional convolutional layers, the first LSTM layer, the second LSTM layer, and the output layer;

[0070] The data of the input layer is [N, t, k], where N represents the sample size of the reservoir water level data, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station within t consecutive days; there are three two-dimensional convolution layers, denoted as C1, C2, and C3, respectively, where the convolution layer is expressed as: C = f(wx+b), where f is the activation function, and all three layers are ReLU activation functions. In order to match the two-dimensional convolution operation, the input layer is expanded by one dimension, and the expanded input shape is [N, t, k, 1]. By setting the size of the convolution kernel and the padding operation, the output after the three-layer two-dimensional convolution layer is further obtained, and its output dimension is [N, m, k, 256], where m represents the value after the three-layer convolution, which includes the time change law of the water level within t days; the number of output channels is 256, which integrates the water level change data between k water level stations;

[0071] The output dimensions after the three convolution layers are combined to obtain an output shape of [N, m*k*256] to match the input of the first LSTM layer. The output shape after the first LSTM layer is [N, 256], where the number of neurons in the middle hidden layer is 256. Finally, the output is sent to the second LSTM layer and the output shape is [N, 512], where the number of output neurons is 512.

[0072] The output layer is a fully connected layer that receives the output from the second LSTM layer as its input and obtains an output of shape [N, 5], where 5 represents the daily average water level data of the five downstream water level stations in the next day;

[0073] The loss function is:

[0074]

[0075] Among them, Y obs is the actual reservoir water level, Y pre is the predicted value of the above neural network model, and N represents the total number of samples.

[0076] In steps 4 and 5, the Adam optimizer based on the gradient descent strategy is used to optimize the weight parameters during the network training process. The batch processing parameter batch_size is set to 200, the number of training rounds is set to 100, the learning rate is set to 0.0001, and the Dropout regularization method is used to reduce network overfitting. After the network model is built based on the training set, the network hyperparameters are adjusted on the validation set. The network weight parameters of the model are determined through multiple verifications, and the M optimal models with the highest prediction accuracy during the verification process are saved.

[0077] In step 6, the Epoch-Wise Bagging ensemble learning strategy includes: building a convolutional neural network and a long short-term memory hybrid model based on the training set, adjusting the weight parameters of the network through the validation set and saving M optimal accuracy models as base learners, and then selecting these M optimal base learners for ensemble learning testing and outputting the prediction results of the M optimal base learners.

[0078] Example

[0079] like Figure 3 As shown, the technical solution of the present invention is further illustrated by taking the water level station of Dongjiang Reservoir as an example, specifically:

[0080] The first step was to obtain hydrological time series data from seven water level stations spatially adjacent to the Dongjiang Reservoir from 1956 to 2000 as water level data samples. After data normalization, the corresponding training, validation, and test sets were obtained. The data samples from 1956 to 1990 served as the training set, the data samples from 1990 to 1995 served as the validation set, and the data samples from 1995 to 2000 served as the test set.

[0081] The second step is to construct a hybrid model based on a convolutional neural network and a long short-term memory network. The reservoir water level prediction model specifically consists of a network input layer, three two-dimensional convolutional layers, the first LSTM layer, the second LSTM layer, and an output layer. Water level data from seven stations—Waizhou Station on the Ganjiang River, Lijiadu Station on the Fuhe River, Meigang Station on the Xinjiang River, Dufengkeng Station on the Raohe River, Hushan Station, Wanjiabu Station on the Xiushui River, and Hankou Station—are used for the joint prediction. This data is fed into the water level prediction model, which then predicts the water levels of five representative stations—Hukou Station, Xingzi Station, Duchang Station, Wucheng Station, and Kangshan Station—for the next day. To improve the prediction accuracy of the reservoir water level model, the minimum unit is daily. In the actual test scenario, t is set to 7, representing the water level values at each station for seven consecutive days. The shapes of the input and output are [N, 7, 7, 1] and [N, 5], respectively.

[0082] Step 3, Figure 4Figure 1 is a diagram of the training process. The network is trained using the MSE (root square error) loss function, and the Adam optimizer is used to optimize the weight parameters. The batch processing parameters batch_size are set to 200, the epoch is set to 100, and the learning rate is set to 0.0001. Dropout regularization is used to prevent overfitting of the network.

[0083] In the fourth step, based on the Epoch-Wise Bagging ensemble learning strategy, a weighted network model was used to predict the spatial distribution of water levels at the five representative stations within one day of the Dongjiang Reservoir in real time, according to the real-time water inflow of the seven water level stations. Furthermore, the network model was compared with other water level prediction models. Figure 5 Schematic diagram comparing the prediction results of the water level prediction model of this embodiment with those of other models.

[0084] In summary, this paper utilizes a hybrid model of convolutional neural networks and long-short-term memory networks, based on an Epoch-Wise Bagging ensemble learning strategy, to effectively capture water level trends at various water level stations. Firstly, a two-dimensional convolutional neural network is used to capture the correlation between water level values at spatially distinct stations, thereby improving the model's information extraction capabilities. Secondly, a long-short-term memory network is used to learn the temporal dependencies and changing trends of reservoir water levels, thereby improving the accuracy of the model's water level predictions.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit it. Any changes made based on the technical solution in accordance with the technical ideas proposed by the present invention fall within the scope of protection of the present invention.

Claims

1. A neural network reservoir water level prediction method based on ensemble learning, characterized in that: The following steps are involved: Step 1: Collect historical water level data of each adjacent water level station in the reservoir space within a preset time period to form reservoir water level sample data; Step 2: Perform numerical transformation on the reservoir water level sample data in step 1, i.e., data normalization, to obtain a standard data set, and divide the standard data set into a training set, a validation set, and a test set; Step 3: Build a hybrid model based on convolutional neural network and long short-term memory network, and define the model loss function; Step 4: Use the training set in step 2 to train the neural network model built in step 3, and use the gradient descent strategy to optimize the loss function of step 3; Step 5: Use the validation set in step 2 to adjust the parameters of the neural network model, determine the network weight parameters of the model through multiple validations, and save the M optimal models with the highest prediction accuracy during the validation process; Step 6: According to the Epoch-Wise Bagging ensemble learning strategy, select the M best network models verified by the validation set in step 5, and test these M best network models using the test set in step 2; Step 7: Input the real-time water inflow of each water level station into M optimal network models, and calculate the M output results corresponding to these M optimal network models using the weighted average method to obtain the final model output to predict the real-time water level spatial distribution of the reservoir within the next 1 day.

2. The method for predicting reservoir water level based on neural network using ensemble learning according to claim 1, characterized in that: In step 1, the water level historical data are collected and formed into an input sample data set for predicting the reservoir water level, and its shape is expressed as [N, t, k], where N represents the total sample size of the input data of the reservoir water level, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station for t consecutive days.

3. The neural network reservoir water level prediction method based on ensemble learning according to claim 2 is characterized in that: The data in step 2 is normalized to: Among them, X cur are the values of t and k that need to be normalized; X min are the minimum values of t and k respectively; X max are the maximum values of t and k respectively; The normalized input data samples are divided into training set, validation set and test set in the ratio of 8:1:

1.

4. The method for predicting reservoir water level based on neural network ensemble learning according to claim 1, characterized in that: In step 3, the convolutional neural network and long short-term memory hybrid model specifically includes the following layers, in order: input layer, three two-dimensional convolutional layers, the first LSTM layer, the second LSTM layer, and the output layer; The data of the input layer is [N, t, k], where N represents the sample size of the reservoir water level data, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station within t consecutive days; there are three two-dimensional convolution layers, denoted as C1, C2, and C3, respectively, where the convolution layer is expressed as: C = f(wx+b), where f is the activation function, and all three layers are ReLU activation functions. In order to match the two-dimensional convolution operation, the input layer is expanded by one dimension, and the expanded input shape is [N, t, k, 1]. By setting the size of the convolution kernel and the padding operation, the output after the three-layer two-dimensional convolution layer is further obtained, and its output dimension is [N, m, k, 256], where m represents the value after the three-layer convolution, which includes the time change law of the water level within t days; the number of output channels is 256, which integrates the water level change data between k water level stations; The output dimensions after the three convolution layers are combined to obtain an output shape of [N, m*k*256] to match the input of the first LSTM layer. The output shape after the first LSTM layer is [N, 256], where the number of neurons in the middle hidden layer is 256. Finally, the output is sent to the second LSTM layer and the output shape is [N, 512], where the number of output neurons is 512. The output layer is a fully connected layer that receives the output from the second LSTM layer as its input and obtains an output of shape [N, 5], where 5 represents the daily average water level data of the five downstream water level stations in the next day; The loss function is: Among them, Y obs is the actual reservoir water level, Y pre is the predicted value of the above neural network model, and N represents the total number of samples.

5. The method for predicting reservoir water level based on neural network using ensemble learning according to claim 4, characterized in that: In steps 4 and 5, the Adam optimizer based on the gradient descent strategy is used to optimize the weight parameters during the network training process. The batch processing parameter batch_size is set to 200, the number of training rounds is set to 100, the learning rate is set to 0.0001, and the Dropout regularization method is used to reduce network overfitting. After the network model is built based on the training set, the network hyperparameters are adjusted on the validation set. The network weight parameters of the model are determined through multiple verifications, and the M optimal models with the highest prediction accuracy during the verification process are saved.

6. The method for predicting reservoir water level based on neural network using ensemble learning according to claim 5, characterized in that: In step 6, the Epoch-Wise Bagging ensemble learning strategy includes: building a convolutional neural network and a long short-term memory hybrid model based on the training set, adjusting the weight parameters of the network through the validation set and saving M optimal accuracy models as base learners, and then selecting these M optimal base learners for ensemble learning testing and outputting the prediction results of the M optimal base learners.

7. A neural network reservoir water level prediction system based on ensemble learning, characterized in that: The system includes the following steps: Data collection module: collects historical water level data of each adjacent water level station in the reservoir space within a preset time period to form reservoir water level sample data; Data processing module: Perform numerical transformation on reservoir water level sample data, i.e. data normalization, to obtain a standard data set, and divide the standard data set into training set, validation set and test set; Neural network building module: Build a hybrid model based on convolutional neural network and long short-term memory network, and define the model loss function; Neural network training module: Use the training set to train the constructed neural network model and use the gradient descent strategy to optimize the loss function; Neural network optimization module: Use the validation set to adjust the parameters of the neural network model, determine the network weight parameters of the model through multiple validations, and save the M optimal models with the highest prediction accuracy during the validation process; Optimal network model selection module: Based on the Epoch-Wise Bagging ensemble learning strategy, select M optimal network models verified by the validation set, and use the test set to test these M optimal network models; Water level distribution prediction module: The real-time water inflow of each water level station is input into M optimal network models, and the M output results corresponding to these M optimal network models are calculated by weighted averaging to obtain the final model output to predict the real-time water level spatial distribution of the reservoir within the next 1 day.

8. The neural network reservoir water level prediction system based on ensemble learning according to claim 7 is characterized in that: In the data collection module, the water level historical data is collected and formed into an input sample data set for predicting the reservoir water level, and its shape is expressed as [N, t, k], where N represents the total sample size of the input data of the reservoir water level, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station for t consecutive days.

9. The neural network reservoir water level prediction system based on ensemble learning according to claim 8, characterized in that: In the data processing module, the data is normalized as follows: Among them, X cur are the values of t and k that need to be normalized; X min are the minimum values of t and k respectively; X max are the maximum values of t and k respectively; The normalized input data samples are divided into training set, validation set and test set in the ratio of 8:1:

1.

10. The neural network reservoir water level prediction system based on ensemble learning according to claim 9, characterized in that: In the neural network building module, the convolutional neural network and long short-term memory hybrid model specifically includes the following layers, in order: input layer, three two-dimensional convolutional layers, the first LSTM layer, the second LSTM layer, and the output layer; The data of the input layer is [N, t, k], where N represents the sample size of the reservoir water level data, k represents the number of adjacent water level stations in the reservoir space, with each day as the minimum unit, and t represents the water level value of each water level station within t consecutive days; there are three two-dimensional convolution layers, denoted as C1, C2, and C3, respectively, where the convolution layer is expressed as: C = f(wx+b), where f is the activation function, and all three layers are ReLU activation functions. In order to match the two-dimensional convolution operation, the input layer is expanded by one dimension, and the expanded input shape is [N, t, k, 1]. By setting the size of the convolution kernel and the padding operation, the output after the three-layer two-dimensional convolution layer is further obtained, and its output dimension is [N, m, k, 256], where m represents the value after the three-layer convolution, which includes the time change law of the water level within t days; the number of output channels is 256, which integrates the water level change data between k water level stations; The output dimensions after the three convolution layers are combined to obtain an output shape of [N, m*k*256] to match the input of the first LSTM layer. The output shape after the first LSTM layer is [N, 256], where the number of neurons in the middle hidden layer is 256. Finally, the output is sent to the second LSTM layer and the output shape is [N, 512], where the number of output neurons is 512. The output layer is a fully connected layer that receives the output from the second LSTM layer as its input and obtains an output of shape [N, 5], where 5 represents the daily average water level data of the five downstream water level stations in the next day; The loss function is: Among them, Y obs is the actual reservoir water level, Y pre is the predicted value of the above neural network model, and N represents the total number of samples.

Images