Reservoir level prediction method based on graph convolutional neural network

Abstract

The invention belongs to the field of hydrological forecasting, and discloses a reservoir level prediction method based on a graph convolutional neural network (GCN). The method aims at solving the reservoir level prediction problem under the rainfall condition, and effective support is provided for reservoir dispatching. Rainfall image data are trained mainly through a graph convolutional neural network (GCN), and a bidirectional long-short term memory network (BiLSTM) is combined to extract time sequence features of the reservoir water level, so that a model capable of predicting the reservoir water level through rainfall images is constructed. Specifically, the GCN is used for extracting spatio-temporal features of a rainfall image and fusing the features by using a multi-head attention mechanism. In BiLSTM, an input time series data sequence is processed through LSTM networks in two directions, so that past and future information is captured at the same time. By combining the hidden states in the two directions, the model can understand the internal law of the time series data more deeply, so that the prediction precision of the reservoir level is improved.

Description

Technical Field

[0001] The present invention belongs to the field of hydrological forecasting, and in particular relates to a reservoir water level prediction method based on a graph convolutional neural network. Technical Background

[0002] Reservoir scheduling is closely related to reservoir water level, which in turn is closely related to rainfall. Therefore, the study of reservoir scheduling can first study the rainfall problem. Reservoir scheduling is affected by many factors, including flood control, power generation, and farmland irrigation. Among them, flood control is the factor that causes the greatest pressure on reservoir scheduling, and floods are often caused by extreme rainfall. Therefore, when faced with extreme rainfall, reservoirs need to make rapid adjustments to reduce the risk of dam breaches.

[0003] Reservoir water level is a basis for reservoir operation, and the change of reservoir water level is closely related to rainfall. Analyzing past rainfall data can accurately grasp the change of reservoir water level, which is crucial for reservoir prediction and operation. On the one hand, with the continuous improvement of hydrological science and technology and the maturity of water conservancy information technology, the hydrological industry is undergoing a transformation from traditional hydrology to modern hydrology. Hydrological data acquisition has gradually evolved from manual recording to automatic acquisition by hydrological stations, accumulating a large amount of rich historical data on water and rainfall conditions, providing new technical means and computing power support for reservoir water level prediction applications. Graph convolutional neural networks can effectively capture the spatial and temporal dependencies in hydrological data, use graph structures to model basin nodes, combine basin rainfall images, and use bidirectional long short-term memory networks (BiLSTM) to extract the time series characteristics of reservoir water levels. Finally, a model that can extract the complex relationship between rainfall images and reservoir water level changes can be obtained, which can effectively improve the accuracy of reservoir water level prediction. Summary of the invention

[0004] The purpose of this invention is to provide a method for predicting reservoir water level data based on graph convolutional neural network. The graph convolutional neural network is used to extract spatiotemporal features and the bidirectional long short-term memory network is used to extract temporal features, providing technical support for the prediction of reservoir water levels and effective support for reservoir scheduling.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for predicting reservoir water level based on graph convolutional neural network, characterized by comprising the following steps:

[0007] Step 1: Collect rainfall image data and reservoir water level data of the target basin for in-depth analysis and modeling;

[0008] Step 2: preprocess the rainfall image data and reservoir water level data to ensure the consistency of data quality;

[0009] Step 3: Align the rainfall image data and reservoir water level data to ensure the consistency of their timestamps and data types for accurate comparison and analysis;

[0010] Step 4: Use a graph convolutional neural network (GCN) to train rainfall image data, and combine it with a bidirectional long short-term memory network (BiLSTM) to extract the time series characteristics of reservoir water levels, thereby building a model that can predict reservoir water levels through rainfall images;

[0011] Step 5: Input the rainfall image data into the trained prediction model to predict the reservoir water level data;

[0012] Furthermore, the step 1 includes the following steps:

[0013] Sub-step 1.1, search for rainfall image data and reservoir water level data from the website of the Ministry of Water Resources of the target basin and the website of the provincial and municipal water resources department to which the basin belongs;

[0014] Sub-step 1.2: After obtaining the initial data from the website of the Ministry of Water Resources and the online website of the provincial and municipal water resources department to which the basin belongs, it is necessary to filter out valuable data according to the needs, such as rainfall image data and reservoir water level data of the specified basin;

[0015] Furthermore, the step 2 includes the following steps:

[0016] Sub-step 2.1, check the collected rainfall image data and reservoir water level data to confirm whether there is any missing data, and appropriately supplement or delete the missing data according to the specific situation;

[0017] Sub-step 2.2, pre-processing the rainfall image data to ensure that the size of each rainfall image is consistent so as to facilitate batch processing;

[0018] Furthermore, the step 3 includes the following steps:

[0019] Sub-step 3.1, unify the time units of rainfall image data and reservoir water level data, and convert the string timestamp into a standard date and time format;

[0020] Sub-step 3.2, ensuring that the data types of all reservoir water level data are consistent for subsequent analysis and processing;

[0021] Furthermore, the step 4 includes the following steps:

[0022] Sub-step 4.1, divide the corresponding rainfall image data and reservoir water level data into training set, validation set and test set in a ratio of 7:1.5:1.5;

[0023] In sub-step 4.2, the graph convolutional neural network is divided into five layers: data preprocessing layer, feature encoder layer, feature decoder layer, water level prediction layer, and output layer. The data preprocessing layer is mainly responsible for preprocessing the hydrological data of the target basin and analyzing the correlation between historical data and future data. The feature encoder layer and the feature decoder layer are responsible for mining and fusing hydrological data features through the graph convolutional neural network (GCN). The water level prediction layer mines the time series hydrological data after feature fusion based on BiLSTM to predict the reservoir water level data at future times.

[0024] Sub-step 4.3, select the rainfall image data that affects the change of reservoir water level data as the key factor. By calculating the correlation coefficient between the historical rainfall image data and the reservoir water level data at the future time, and analyzing the cross-correlation diagram between the historical rainfall image data and the reservoir water level data at the future time, the data volume of the influencing factor is determined. The Pearson correlation coefficient is used to measure the linear relationship strength between the two time series of rainfall image data and reservoir water level data. The calculation formula is as follows:

[0025]

[0026] Where xi and yi are the i-th data of sequences X and Y respectively; and represent the mean values ​​of series X and Y respectively.

[0027] Sub-step 4.4, training the prediction model, selecting appropriate data from the training set for model training, using the encoder-decoder structure, extracting the spatiotemporal features of the rainfall image through the convolutional neural network (GCN), using the multi-head attention mechanism to fuse these features, and finally using BiLSTM to predict the reservoir water level data. During the training process, the correlation coefficient between the rainfall image data at the historical moment and the reservoir water level data at the future moment is calculated, and the cross-correlation diagram between the rainfall image data at the historical moment and the reservoir water level data at the future moment is analyzed to determine the amount of data affecting the factors. The cross-correlation function is as follows:

[0028] R=E(X s *Y t )

[0029] Where, X s represents the value of the time series X at time s, Y t It represents the value of time series Y at time t, and R represents the correlation between the two.

[0030] The data that has passed through the encoder-decoder and uses the multi-head attention mechanism to fuse features is used as the input of the bidirectional long short-term memory network (BiLSTM) for the final reservoir water level prediction. In BiLSTM, the input time series data sequence is processed by the LSTM network in two directions to capture past and future information. By merging the hidden states in these two directions, the model can better understand the inherent laws of the time series data, thereby improving the prediction accuracy. The prediction function is as follows:

[0031] y′=f BiLSTM (X′)

[0032] Among them, X' is the fused rainfall image data, and y' is the predicted value of reservoir water level.

[0033] In sub-step 4.5, the prediction model based on graph convolutional neural network (GCN) is trained by calculating the loss (Loss) between the predicted value y' and the true value y. The loss function is composed of the rainfall image data error (PE) and the mean square error (MSE).

[0034] The calculation formula of rainfall image data error is as follows:

[0035] PE=y max -y max

[0036] The formula for the mean square error is as follows:

[0037]

[0038] The loss function calculation formula is as follows:

[0039] Loss=PE+MSE

[0040] Furthermore, the step 5 includes the following steps:

[0041] Sub-step 5.1, a denormalization process is performed on the prediction output of the prediction model to convert the prediction value back to the original data range, enhance the interpretability and intuitiveness of the result, more accurately verify the prediction performance of the model, and finally output the prediction result of the reservoir water level data.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention proposes a reservoir water level prediction method based on graph convolutional neural network. The graph convolutional neural network can effectively capture the spatial and temporal dependencies in rainfall image data, and further extract the temporal relationship of reservoir water level data changes in combination with a bidirectional long short-term memory network, thereby deeply exploring the impact of rainfall image data on reservoir water level, so that reservoir water level prediction can be achieved based on rainfall image data. This method can timely predict the subsequent changes in reservoir water levels during rainfall, providing a scientific basis for formulating response measures. This not only helps to optimize reservoir scheduling strategies, but also effectively supports flood prevention work, and provides strong technical support and decision-making support for water resources management. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] Figure 1 It is a basic implementation flow chart of the present invention.

[0046] Figure 2 Graph of the construction of a graph convolutional neural network.

[0047] Figure 3 This is the implementation diagram of the bidirectional long short-term memory network.

[0048] Figure 4 Construct a graph for the predictive model of the present invention.

[0049] Figure 5 Flowchart for the implementation of model prediction. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.

[0051] A reservoir water level prediction method based on graph convolutional neural network in this embodiment is as follows: Figure 1 The following steps are shown:

[0052] Step 1: Collect rainfall image data and reservoir water level data of the target basin for in-depth analysis and modeling;

[0053] Furthermore, the step 1 includes the following steps:

[0054] Sub-step 1.1, search for rainfall image data and reservoir water level data from the website of the Ministry of Water Resources of the target basin and the website of the provincial and municipal water resources department to which the basin belongs;

[0055] Sub-step 1.2, determine the range of rainfall image data and reservoir water level data that need to be obtained on the website of the Ministry of Water Resources of the target basin and the online website of the provincial and municipal water resources department to which the basin belongs, and screen the initial data within the range to select valuable data according to the needs.

[0056] Step 2: preprocess the rainfall image data and reservoir water level data to ensure the consistency of data quality;

[0057] Furthermore, the step 2 includes the following steps:

[0058] Sub-step 2.1, after collecting the data, check the integrity of the collected rainfall image data and reservoir water level data to confirm whether there is any missing data, and appropriately supplement or delete the missing data according to the specific situation;

[0059] Sub-step 2.2, pre-processing the rainfall image data to ensure that the size of each rainfall image is consistent so as to facilitate batch processing;

[0060] Step 3: Align the rainfall image data and reservoir water level data to ensure the consistency of their timestamps and data types for accurate comparison and analysis;

[0061] Furthermore, the step 3 includes the following steps:

[0062] Sub-step 3.1, unify the time units of rainfall image data and reservoir water level data, convert the string timestamp into a standard date and time format, so that the time of rainfall image data can correspond to the time of reservoir water level data;

[0063] Sub-step 3.2, ensuring that the data types of all reservoir water level data are consistent for subsequent analysis and processing;

[0064] Step 4: Use a graph convolutional neural network (GCN) to train rainfall image data, and combine it with a bidirectional long short-term memory network (BiLSTM) to extract the time series characteristics of reservoir water levels, thereby building a model that can predict reservoir water levels through rainfall images;

[0065] Furthermore, the step 4 includes the following steps:

[0066] Sub-step 4.1, divide the corresponding rainfall image data and reservoir water level data into training set, validation set and test set in the ratio of 7:1.5:1.5:

[0067] Sub-step 4.2, the graph convolutional neural network is as follows Figure 2As shown, it is divided into five layers: data preprocessing layer, feature encoder layer, feature decoder layer, water level prediction layer, and output layer. The data preprocessing layer is mainly responsible for preprocessing the hydrological data of the target basin and analyzing the correlation between historical data and future data. The feature encoder layer and the feature decoder layer are responsible for mining and fusing hydrological data features through graph convolutional neural network (GCN). The water level prediction layer mines the time series hydrological data after feature fusion based on BiLSTM to predict the reservoir water level data at future times.

[0068] Sub-step 4.3, select the rainfall image data that affects the change of reservoir water level data as the key factor. By calculating the correlation coefficient between the historical rainfall image data and the reservoir water level data at the future time, and analyzing the cross-correlation diagram between the historical rainfall image data and the reservoir water level data at the future time, the data volume of the influencing factor is determined. The Pearson correlation coefficient is used to measure the linear relationship strength between the two time series of rainfall image data and reservoir water level data. The calculation formula is as follows:

[0069]

[0070] In the formula, x i and i are the i-th data of sequences X and Y respectively; and represent the mean values ​​of series X and Y respectively.

[0071] Substep 4.4, training the prediction model, select appropriate data from the training set for model training, use the encoder-decoder structure, extract the spatiotemporal features of the rainfall image through the convolutional neural network (GCN), use the multi-head attention mechanism to fuse these features, and finally use Figure 3 The BiLSTM shown in the figure predicts the reservoir water level data. During the training process, the correlation coefficient between the rainfall image data at the historical moment and the reservoir water level data at the future moment is calculated, and the cross-correlation diagram between the rainfall image data at the historical moment and the reservoir water level data at the future moment is analyzed to determine the data volume of the influencing factors. The cross-correlation function is as follows:

[0072] R=E(X s *Y t )

[0073] Where, X s represents the value of the time series X at time s, Y t It represents the value of time series Y at time t, and R represents the correlation between the two.

[0074] The data that has passed through the encoder-decoder and uses the multi-head attention mechanism to fuse features is used as the input of the bidirectional long short-term memory network (BiLSTM) for the final reservoir water level prediction. In BiLSTM, the input time series data sequence is processed by the LSTM network in two directions to capture past and future information. By merging the hidden states in these two directions, the model can better understand the inherent laws of the time series data, thereby improving the prediction accuracy. The prediction function is as follows:

[0075] y′=f BiLSTM (X′)

[0076] Among them, X' is the fused rainfall image data, and y' is the predicted value of reservoir water level.

[0077] In sub-step 4.5, the prediction model based on graph convolutional neural network (GCN) is trained by calculating the loss (Loss) between the predicted value y' and the true value y. The loss function is composed of the rainfall image data error (PE) and the mean square error (MSE).

[0078] The calculation formula of rainfall image data error is as follows:

[0079] PE=y max -y max

[0080] The formula for the mean square error is as follows:

[0081]

[0082] The loss function calculation formula is as follows:

[0083] Loss=PE+MSE

[0084] Step 5: Input the rainfall image data into the trained prediction model to predict the reservoir water level data;

[0085] Furthermore, the step 5 includes the following steps:

[0086] Sub-step 5.2, a denormalization process is performed on the prediction output of the prediction model to convert the prediction value back to the original data range, enhance the interpretability and intuitiveness of the result, more accurately verify the prediction performance of the model, and finally output the prediction result of the reservoir water level data.

Claims

1. A method for predicting reservoir water level based on graph convolutional neural network, comprising the following steps: Step 1: Collect rainfall image data and reservoir water level data of the target basin for in-depth analysis and modeling; Step 2: preprocess the rainfall image data and reservoir water level data to ensure the consistency of data quality; Step 3: Align the rainfall image data and reservoir water level data to ensure the consistency of their timestamps and data types for accurate comparison and analysis; Step 4: Use a graph convolutional neural network (GCN) to train rainfall image data, and combine it with a bidirectional long short-term memory network (BiLSTM) to extract the time series characteristics of reservoir water levels, thereby building a model that can predict reservoir water levels through rainfall images; Step 5: Input the rainfall image data into the trained prediction model to predict the reservoir water level data.

2. The method for predicting reservoir water level based on graph convolutional neural network according to claim 1, characterized in that: Step 1 contains the following sub-steps: Sub-step 1.1, search for rainfall image data and reservoir water level data from the website of the Ministry of Water Resources of the target basin and the website of the provincial and municipal water resources department to which the basin belongs; Sub-step 1.2: After obtaining the initial data from the website of the Ministry of Water Resources and the online website of the provincial and municipal water resources department to which the basin belongs, it is necessary to filter out valuable data according to needs, such as rainfall image data and reservoir water level data of the specified basin.

3. The method for predicting reservoir water level based on graph convolutional neural network according to claim 2 is characterized in that: Step 2 contains the following sub-steps: Sub-step 2.1, check the collected rainfall image data and reservoir water level data to confirm whether there is any missing data, and appropriately supplement or delete the missing data according to the specific situation; Sub-step 2.2, pre-processing the rainfall image data to ensure that the size of each rainfall image is consistent so as to facilitate batch processing.

4. The method for predicting reservoir water level based on graph convolutional neural network according to claim 3 is characterized in that: Step 3 contains the following sub-steps: Sub-step 3.1, unify the time units of rainfall image data and reservoir water level data, and convert the string timestamp into a standard date and time format; Sub-step 3.2, ensure that the data types of all reservoir water level data are consistent to facilitate subsequent analysis and processing.

5. The method for predicting reservoir water level based on graph convolutional neural network according to claim 4 is characterized in that: Step 4 contains the following sub-steps: Sub-step 4.1, divide the corresponding rainfall image data and reservoir water level data into training set, validation set and test set in a ratio of 7:1.5:1.5; In sub-step 4.2, the graph convolutional neural network is divided into five layers: data preprocessing layer, feature encoder layer, feature decoder layer, water level prediction layer, and output layer. The data preprocessing layer is mainly responsible for preprocessing the hydrological data of the target basin and analyzing the correlation between historical data and future data. The feature encoder layer and the feature decoder layer are responsible for mining and fusing hydrological data features through the graph convolutional neural network (GCN). The water level prediction layer mines the time series hydrological data after feature fusion based on BiLSTM to predict the reservoir water level data at future times. Sub-step 4.3, select the rainfall image data that affects the change of reservoir water level data as the key factor. By calculating the correlation coefficient between the historical rainfall image data and the reservoir water level data at the future time, and analyzing the cross-correlation diagram between the historical rainfall image data and the reservoir water level data at the future time, the data volume of the influencing factor is determined. The Pearson correlation coefficient is used to measure the linear relationship strength between the two time series of rainfall image data and reservoir water level data. The calculation formula is as follows: In the formula, x i and i are the i-th data of sequences X and Y respectively; and represent the mean values ​​of series X and Y respectively. Sub-step 4.4, training the prediction model, selecting appropriate data from the training set for model training, using the encoder-decoder structure, extracting the spatiotemporal features of the rainfall image through the convolutional neural network (GCN), using the multi-head attention mechanism to fuse these features, and finally using BiLSTM to predict the reservoir water level data. During the training process, the correlation coefficient between the rainfall image data at the historical moment and the reservoir water level data at the future moment is calculated, and the cross-correlation diagram between the rainfall image data at the historical moment and the reservoir water level data at the future moment is analyzed to determine the amount of data affecting the factors. The cross-correlation function is as follows: R=E(X s *Y t ) Where, X s represents the value of the time series X at time s, Y t It represents the value of time series Y at time t, and R represents the correlation between the two. The data that has passed through the encoder-decoder and uses the multi-head attention mechanism to fuse features is used as the input of the bidirectional long short-term memory network (BiLSTM) for the final reservoir water level prediction. In BiLSTM, the input time series data sequence is processed by the LSTM network in two directions to capture past and future information. By merging the hidden states in these two directions, the model can better understand the inherent laws of the time series data, thereby improving the prediction accuracy. The prediction function is as follows: y′=f BiLSTM (X′) Among them, X' is the fused rainfall image data, and y' is the predicted value of reservoir water level. In sub-step 4.5, the prediction model based on graph convolutional neural network (GCN) is trained by calculating the loss (Loss) between the predicted value y' and the true value y. The loss function is composed of the rainfall image data error (PE) and the mean square error (MSE). The calculation formula of rainfall image data error is as follows: PE=y max -y max ′ The formula for the mean square error is as follows: The loss function calculation formula is as follows: Loss=PE+MSE.

6. The method for predicting reservoir water level based on graph convolutional neural network according to claim 5, characterized in that: Step 5 contains the following sub-steps: Sub-step 5.1, a denormalization process is performed on the prediction output of the prediction model to convert the prediction value back to the original data range, enhance the interpretability and intuitiveness of the result, more accurately verify the prediction performance of the model, and finally output the prediction result of the reservoir water level data.

Images