Water level prediction method based on deep learning networked watershed inundation model

Abstract

The present invention relates to a water level prediction method based on a deep learning networked watershed inundation model, comprising the following steps: step S1: obtaining rainwater condition data of each measuring point in the watershed and preprocessing; step S2: performing feature extraction and superposition operations on the preprocessed rainwater condition data; step S3: performing training and fitting based on an LSTM model according to the extracted and superimposed features; step S4: using PSO to search for the optimal parameters of the LSTM model after training in step S3, and obtaining a deep learning networked watershed inundation model; step S5: performing watershed water level prediction based on the deep learning networked watershed inundation model. The present invention can realize the inheritance of monitoring data at different measuring points, effectively avoid the waste of computing resources caused by repeated calculations, and improve prediction efficiency.

Description

Technical Field

[0001] The present invention relates to the field of water level prediction, and in particular to a water level prediction method based on a deep learning networked watershed inundation model. Background Art

[0002] With the rapid development of artificial intelligence technology and the gradual coverage of river basin rainfall monitoring systems, the use of machine learning to predict river water levels has been applied to flood prevention and early warning systems for rivers with flood risks.

[0003] The basin inundation model is based on the water level prediction of multiple measuring points in the basin. The traditional method builds a model for each water level prediction point. Each model is trained and fitted using the relevant data of the water level prediction point in the basin data. Therefore, this technical solution has three disadvantages:

[0004] 1. Too many models. The number of models is equal to the number of prediction points. When there are many prediction points, the workload of model training and deployment is high, and the operation and maintenance cost is high. This problem is common in medium and large basins.

[0005] 2. Repeated feature extraction causes a waste of computing resources. In the water level prediction models at different measuring points, features appear repeatedly at a very high frequency. Each model will extract these repeated features separately, which greatly wastes computing resources.

[0006] 3. Especially when the amount of data in the early stage is relatively small and the later model needs to be continuously iterated, the operation and maintenance cost is high, which will result in a large iteration cost. Summary of the invention

[0007] In view of this, the purpose of the present invention is to provide a water level prediction method based on a deep learning networked watershed inundation model, which can realize the inheritance of monitoring data from different measuring points, effectively avoid the waste of computing resources caused by repeated calculations, and improve prediction efficiency.

[0008] To achieve the above object, the present invention adopts the following technical solution:

[0009] A water level prediction method based on a deep learning networked watershed inundation model comprises the following steps:

[0010] Step S1: Obtain the rainfall data of each measuring point in the basin and pre-process it;

[0011] Step S2: performing feature extraction and overlay operations on the pre-processed rainfall data;

[0012] Step S3: Perform training and fitting based on the LSTM model according to the extracted and superimposed features;

[0013] Step S4: Use PSO to search for the optimal parameters of the LSTM model trained in step S3 to obtain a deep learning-based networked watershed inundation model;

[0014] Step S5: Based on the deep learning networked watershed inundation model, basin water level prediction is performed.

[0015] Furthermore, the step S1 is specifically:

[0016] Step S11: Obtain the rainfall data of each measuring point in the watershed, reduce the rainfall data to the same precision, and then use the acquisition time as a keyword to connect the data;

[0017] Step S12: The connected data is sorted by acquisition time.

[0018] Furthermore, the rain and water situation data include rainfall and water level data, and the reservoir measuring points additionally include flood discharge data.

[0019] Furthermore, the step S2 is specifically as follows:

[0020] Step S21: Select the data of the upstream measuring point of the basin as the initial input, use conv1D to convolve the input, extract the information of the input layer, and then use the fully connected layer to map the convolution layer output to a higher dimensional space. The fully connected layer has two output directions: one is input downward to the model for training, and the other is horizontally superimposed to the next input;

[0021] Step S22: gradually input the measurement point data according to the upstream and downstream relationship of the basin space, and input the rainfall and water level data of one measurement point each time. Use conv1D to extract the input information in the same way as step S21, and map it to the high-dimensional space through the fully connected layer. The calculation formula of the fully connected layer is as follows:

[0022]

[0023] Among them, D j is the jth output, f(x) is the activation function, W ji is the i-th input weight parameter of the j-th output, ε j is the bias parameter of the jth output, and n is the total number of inputs;

[0024] The merge layer is used to merge the fully connected layer output of the previous measurement point. The output of the merge layer is still in two directions: one is input to the LSTM layer to obtain long-term and short-term information, and the other is superimposed on the next input. The feature superposition process is as follows:

[0025]

[0026] In the formula, The input Xi , G i The dataset is convolved and then fully connected to the data matrix vector. in The data matrix vector after convolution and full connection of the sensor's newly added features or data.

[0027] Furthermore, the LSTM model adopts MSE as the loss function. When the number of prediction points is greater than 1, the water level prediction model has multiple groups of MSE. The mean of multiple groups of MSE is taken as the objective function, or the weight of the evaluation index is allocated according to the importance of the prediction point. The weight construction method is shown in the formula:

[0028]

[0029] In the formula, w i is the weight of the MSE of each model, MSE i , are the MSE of the ith model and the weighted average of the MSE of all models, respectively.

[0030] Furthermore, the parameters to be optimized include five hyperparameters: window width, number of neurons in the LSTM layer, Dropout rate, learning rate, and batch size.

[0031] Furthermore, the step S4 is specifically as follows:

[0032] The particle swarm optimization algorithm is used to optimize these five hyperparameters. For each additional water level prediction point, the number of LSTM layers and Dropout layers will be increased by 1 respectively. Therefore, the number of hyperparameters optimized for the final model is 3+2n, where n is the number of water level prediction points.

[0033] MSE is constructed as the objective function of PSO. When the number of prediction points is greater than 1, there are multiple groups of MSE in the water level prediction model. The mean of multiple groups of MSE is taken as the objective function, or the weight of the evaluation index is allocated according to the importance of the prediction point. The weight construction method and its iteration method in the particle swarm are shown in the formula:

[0034]

[0035] Where W(X) is the optimal model hyperparameter combination for PSO, where the optimal hyperparameters include the window width of each model, the number of neurons in the LSTM layer, the Dropout rate, the learning rate, and the batch size. i is the weight of the MSE of each model, MSE i is the MSE of the i-th model;

[0036] According to the optimal parameter combination found by the particle swarm, a networked watershed inundation model based on deep learning is obtained.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The present invention can complete the water level prediction of the entire basin using only one model, which can greatly reduce the workload of model deployment;

[0039] 2. The present invention can realize the inheritance of monitoring data from different measuring points, effectively avoid the waste of computing resources caused by repeated calculations, and improve the prediction efficiency;

[0040] 3. The present invention can be applied to establish a complex watershed monitoring model. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a schematic diagram of a model training process in one embodiment of the present invention;

[0042] Figure 2 It is a schematic diagram of the data extraction and superposition process in one embodiment of the present invention;

[0043] Figure 3 is a schematic diagram of a training and fitting process in one embodiment of the present invention;

[0044] Figure 4 is a PSO flow chart in one embodiment of the present invention;

[0045] Figure 5 It is a schematic diagram of a method according to an embodiment of the present invention;

[0046] Figure 6 is a schematic diagram of a complex watershed according to an embodiment of the present invention;

[0047] Figure 7 This is a watershed inundation model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0049] Please refer to Figure 1-7 The present invention provides a water level prediction method based on a deep learning networked watershed inundation model, comprising the following steps:

[0050] Step S1: Obtain the rainfall data of each measuring point in the basin and pre-process it;

[0051] Step S2: performing feature extraction and overlay operations on the pre-processed rainfall data;

[0052] Step S3: Perform training and fitting based on the LSTM model according to the extracted and superimposed features;

[0053] Step S4: Use PSO to search for the optimal parameters of the LSTM model trained in step S3 to obtain a deep learning-based networked watershed inundation model;

[0054] Step S5: Based on the deep learning networked watershed inundation model, basin water level prediction is performed.

[0055] In this embodiment, the rainfall data is the rainfall and water level data of each measuring point in the basin, and the reservoir measuring point also contains flood discharge data. The collection time of the rainfall data is reduced to the same precision, and then the collection time is used as the keyword to connect the data. The connected data is sorted by the collection time. After sorting, the data set is a set of time series, and its collection time is set to T i Where (i=0,1,2,…,N). The specific data format is as follows:

[0056]

[0057] In this embodiment, data extraction and superposition are performed by Figure 2 Process implementation can be divided into initial input and subsequent input according to its differences.

[0058] Step S21: Select the data of the upstream measuring point of the basin as the initial input, use conv1D to convolve the input, extract the information of the input layer, and then use the fully connected layer to map the convolution layer output to a higher dimensional space. The fully connected layer has two output directions: one is input downward to the model for training, and the other is horizontally superimposed to the next input;

[0059] Step S22: gradually input the measurement point data according to the upstream and downstream relationship of the basin space, and input the rainfall and water level data of one measurement point each time. Use conv1D to extract the input information in the same way as step S21, and map it to the high-dimensional space through the fully connected layer. The calculation formula of the fully connected layer is as follows:

[0060]

[0061] Among them, D j is the jth output, f(x) is the activation function, W ji is the i-th input weight parameter of the j-th output, ε j is the bias parameter of the jth output, and n is the total number of inputs;

[0062] The merge layer is used to merge the fully connected layer output of the previous measurement point. The output of the merge layer is still in two directions: one is input to the LSTM layer to obtain long-term and short-term information, and the other is superimposed on the next input. The feature superposition process is as follows:

[0063]

[0064] In the formula, The input X i , G i The dataset is convolved and then fully connected to the data matrix vector. in The data matrix vector after convolution and full connection of the sensor's newly added features or data.

[0065] In this embodiment, reference Figure 3 , input the extracted and superimposed features into the model, specify the output corresponding to the input, that is, clarify the one or more measurement point data inputs corresponding to each output, and use the LSTM model for training and fitting. The fitting process is as follows Figure 3 As shown; MSE is used as the loss function. When the number of prediction points is greater than 1, the water level prediction model has multiple groups of MSE. The mean of multiple groups of MSE is taken as the objective function, or the weight of the evaluation index is allocated according to the importance of the prediction point. The weight construction method is shown in the formula:

[0066]

[0067] In the formula, w i is the weight of the MSE of each model, MSE i , are the MSE of the ith model and the weighted average of the MSE of all models, respectively.

[0068] In this embodiment, the parameters to be optimized include five hyperparameters: window width, number of neurons in the LSTM layer, Dropout rate, learning rate and batch size.

[0069] The particle swarm optimization algorithm is used to optimize these five hyperparameters. For each additional water level prediction point, the number of LSTM layers and Dropout layers will be increased by 1 respectively. Therefore, the number of hyperparameters optimized for the final model is 3+2n, where n is the number of water level prediction points.

[0070] MSE is constructed as the objective function of PSO. When the number of prediction points is greater than 1, there are multiple groups of MSE in the water level prediction model. The mean of multiple groups of MSE is taken as the objective function, or the weight of the evaluation index is allocated according to the importance of the prediction point. The weight construction method and its iteration method in the particle swarm are shown in the formula:

[0071]

[0072] Where W(X) is the optimal model hyperparameter combination for PSO, where the optimal hyperparameters include the window width of each model, the number of neurons in the LSTM layer, the Dropout rate, the learning rate, and the batch size. i is the weight of the MSE of each model, MSE i is the MSE of the i-th model; the particle swarm optimization process is as follows Figure 4 shown.

[0073] The optimal parameter combination found by the particle swarm is filled into the multi-point water level prediction model training to obtain the multi-point water level prediction model, which is stored in the cloud or locally for subsequent deployment.

[0074] Embodiment 1:

[0075] In this embodiment, a method is proposed to solve the problem of repeated feature extraction in the watershed flooding model. The specific technology is to use a merging function to superimpose the convolved features. The implementation method is as follows:

[0076]

[0077] In the formula, The input X i , G i The dataset is convolved and then fully connected to the data matrix vector. in The data matrix vector after convolution and full connection of the sensor's newly added features or data.

[0078] The present invention proposes a deep learning networked watershed inundation model that can simultaneously predict the water level data of all measuring points in the watershed. The specific implementation method is as follows: Figure 5 As shown, the rainfall data of all measuring points are taken as input. Based on feature extraction and superposition processing, the feature data that has a greater impact on the water level of each measuring point is input into the prediction model corresponding to the measuring point. Each measuring point prediction model outputs the predicted water level of the measuring point, thereby realizing the simultaneous prediction of water levels at multiple measuring points.

[0079] The deep learning networked watershed flooding model constructed by the method of the present invention can be applied to complex watersheds, such as Figure 6 In the complex basin conditions shown in Figure 1, the water level at each measuring point is mainly affected by the upstream measuring point, so the basin inundation model can be deformed as follows: Figure 7 As shown:

[0080] The arrow represents the direction of data transmission, and merge(x,y) represents the merging of the convolution feature of the measurement point x with the convolution feature of the measurement point y.

[0081] In this embodiment, for the evaluation indicators of multiple models, it is proposed to allocate the weights of the evaluation indicators according to the importance of the prediction points, and the weight construction method is shown in the formula:

[0082]

[0083] In the formula, w i is the weight of the MSE of each model, MSE i , are the MSE of the ith model and the weighted average of the MSE of all models, respectively.

[0084] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made according to the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A water level prediction method based on deep learning networked watershed inundation model, characterized in that: The following steps are involved: Step S1: Obtaining the rainfall data of each measuring point in the basin and preprocessing it; the rainfall data includes rainfall and water level data, and the reservoir measuring point additionally includes flood discharge data; Step S2: Perform feature extraction and overlay operations on the pre-processed rainfall data, specifically: Step S21: Select the data of the upstream measuring point of the basin as the initial input, use conv1D to convolve the input, extract the information of the input layer, and then use the fully connected layer to map the convolution layer output to a higher dimensional space. The fully connected layer has two output directions: one is input downward to the model for training, and the other is horizontally superimposed to the next input; Step S22: gradually input the measurement point data according to the upstream and downstream relationship of the basin space, and input the rainfall and water level data of one measurement point each time. Use conv1D to extract the input information in the same way as step S21, and map it to the high-dimensional space through the fully connected layer. The calculation formula of the fully connected layer is as follows: Among them, D j is the jth output, f(x) is the activation function, W ji is the i-th input weight parameter of the j-th output, x i is the i-th input, ε j is the bias parameter of the jth output, and n is the total number of inputs; The Merge layer is used to merge the fully connected layer output of the previous measurement point. The output of the Merge layer is still in two directions: one is input to the LSTM layer to obtain long-term and short-term information, and the other is superimposed on the next input. The feature superposition process is as follows: In the formula, The input X i , G i The dataset is convolved and then fully connected to the data matrix vector. in The data matrix vector after convolution and full connection of the added features or data of the sensor; Step S3: Perform training and fitting based on the LSTM model according to the extracted and superimposed features; Step S4: Use PSO to search for the optimal parameters of the LSTM model trained in step S3, and obtain a deep learning-based networked watershed inundation model, specifically: The particle swarm optimization algorithm is used to optimize these five hyperparameters. For each additional water level prediction point, the number of LSTM layers and Dropout layers will be increased by 1 respectively. Therefore, the number of hyperparameters optimized for the final model is 3+2n, where n is the number of water level prediction points. MSE is constructed as the objective function of PSO. When the number of prediction points is greater than 1, there are multiple groups of MSE in the water level prediction model. The mean of multiple groups of MSE is taken as the objective function, or the weight of the evaluation index is allocated according to the importance of the prediction point. The weight construction method and its iteration method in the particle swarm are shown in the formula: Where W(X) is the optimal model hyperparameter combination for PSO, where the optimal hyperparameters include the window width of each model, the number of neurons in the LSTM layer, the Dropout rate, the learning rate, and the batch size. i is the weight of the MSE of each model, MSE i is the MSE of the i-th model; According to the optimal parameter combination found by the particle swarm, a networked watershed flooding model based on deep learning is obtained; Step S5: Based on the deep learning networked watershed inundation model, basin water level prediction is performed.

2. The water level prediction method based on deep learning networked watershed inundation model according to claim 1 is characterized in that: The step S1 is specifically: Step S11: Obtain the rainfall data of each measuring point in the watershed, reduce the rainfall data to the same precision, and then use the acquisition time as a keyword to connect the data; Step S12: The connected data is sorted by collection time.

3. The water level prediction method based on deep learning networked watershed inundation model according to claim 1 is characterized in that: The LSTM model adopts MSE as the loss function. When the number of prediction points is greater than 1, the water level prediction model has multiple groups of MSE. The mean of multiple groups of MSE is taken as the objective function, or the weight of the evaluation index is allocated according to the importance of the prediction point. The weight construction method is shown in the formula: In the formula, w i is the weight of the MSE of each model, are the MSE of the ith model and the weighted average of the MSE of all models, respectively.

4. The water level prediction method based on deep learning networked watershed inundation model according to claim 1 is characterized in that: The parameters to be optimized include five hyperparameters: window width, number of neurons in the LSTM layer, Dropout rate, learning rate, and batchsize.

Images