A hydrological time series prediction method, system, medium and device

Abstract

The application provides a hydrological time series prediction method, system, medium and equipment, the method comprises the following steps: obtaining a hydrological time series; sequentially performing time series convolution and dimension adjustment on the hydrological time series to obtain a dimension-adjusted hydrological feature; inputting the dimension-adjusted hydrological feature into a pre-trained hydrological prediction model to output a prediction result of the hydrological time series. The application expands the receptive field through time series convolution, thereby expanding the feature extraction range, and the extracted features can cover the information of multiple effective time points; the relative position information is encoded through a rotation matrix, the original data vector is directly rotated, the dependence between adjacent data is excavated, the original data is fully utilized, and no position vector is added, so that the prediction effect is good.

Description

Technical Field

[0001] This invention relates to the field of hydrological data prediction technology, and in particular to a method, system, medium, and equipment for predicting hydrological time series. Background Technology

[0002] Pumping stations are crucial water conservancy infrastructure projects. Accurate water level prediction can scientifically guide dispatchers in water allocation, especially the prediction of water level in the forebay of pumping stations. Hydrological data, such as water levels, are characterized by rapid updates, large volumes, and complex types. Uncovering and analyzing the complex relationships within hydrological data for use in hydrological and flood forecasting has gradually become a key focus.

[0003] Traditional water level prediction is based on physical models of the hydrological environment combined with human experience, which has certain limitations and applicable conditions. With the continuous advancement of artificial intelligence technology, neural networks can be trained from the changing patterns of historical data to develop nonlinear mapping models, which are widely used in water level prediction, including water levels in water diversion projects, lakes, and reservoirs. However, existing water level prediction models suffer from low prediction accuracy, slow training convergence speed, and difficulty in analyzing data based on the correlations between different features. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, system, medium and device for predicting hydrological time series. By expanding the receptive field through temporal convolution, the feature extraction range is expanded. By encoding relative position information through a rotation matrix, the original data vector is directly rotated to uncover the dependencies between adjacent data, making full use of the original data. Without adding position vectors, the model has good prediction performance.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] On one hand, the present invention provides a method for predicting hydrological time series, comprising:

[0007] Obtain hydrological time series;

[0008] The hydrological time series is sequentially subjected to temporal convolution and dimensionality adjustment to obtain the dimensionally adjusted hydrological features.

[0009] The hydrological features adjusted by the dimensions are input into a pre-trained hydrological prediction model, which outputs the prediction results of the hydrological time series.

[0010] The processing steps of the hydrological prediction model include:

[0011] The relative position of the dimension-adjusted hydrological features is encoded using a rotation matrix to obtain hydrological features containing location information.

[0012] An encoder is used to perform sparse correlation calculation and feature compression on hydrological features containing location information to obtain a hydrological feature map.

[0013] A decoder is used to perform masking operations and sparse correlation calculations on hydrological features containing location information to obtain masked hydrological features. The masked hydrological features are then compared with the hydrological feature map to obtain an output feature map.

[0014] The output feature map is predicted using a second fully connected layer to obtain the prediction results of the hydrological time series.

[0015] Optionally, before performing temporal convolution on the hydrological sequence, missing values ​​in the hydrological sequence are filled using a near-distance interpolation method.

[0016] When the missing value of the hydrological sequence is a single value, the average of the two values ​​before and after the missing value is taken as the missing value;

[0017] When the missing values ​​of the hydrological sequence are multi-valued, the formula for the missing values ​​is:

[0018] ;

[0019] In the formula, For the first One missing value. ; The value preceding the missing value; The value following the missing value; The number of missing values; This represents the number of values ​​preceding the missing value.

[0020] Optionally, the hydrological time series can be subjected to temporal convolution using a temporal convolution module to obtain hydrological features;

[0021] The temporal convolution module includes multiple interconnected residual blocks, and the dilation rate of the multiple interconnected residual blocks increases sequentially.

[0022] Optionally, residual connections are introduced into each residual block, and the calculation formula for the hydrological characteristics is as follows:

[0023] ;

[0024] In the formula, Hydrological characteristics; For residual mapping; This is a hydrological time series; It is the first weight; It is the second weight.

[0025] Optionally, the formula for calculating the rotation matrix is:

[0026] ;

[0027] In the formula, The angle is the rotation angle.

[0028] Optionally, the calculation formula for the hydrological features containing location information is as follows:

[0029] ;

[0030] In the formula, For the first A hydrological feature matrix containing location information; For the first Hydrological feature matrix adjusted in each dimension; For matrix The first component; For matrix The second component; For the first Rotation matrix at each position; For the first The rotation angle at each position.

[0031] Optionally, the formula for calculating the sparse correlation is:

[0032] ;

[0033] In the formula, For query matrix; The key matrix; It is the eigenvalue matrix; For the Softmax function; for and KL divergence analysis was used to determine the top activity levels. indivual ; It is the transpose of the key matrix; The dimension of the feature vector.

[0034] Secondly, the present invention provides a hydrological time series prediction system, comprising:

[0035] The data acquisition module is used to acquire hydrological time series data.

[0036] The temporal convolution module is used to perform temporal convolution on hydrological time series.

[0037] The first fully connected layer is used for dimensional adjustment of hydrological time series.

[0038] The encoder is used to perform sparse correlation calculation and feature compression on hydrological features containing location information;

[0039] The decoder is used to perform sparse correlation calculation on the masked hydrological features containing location information, and to perform sparse correlation calculation on the masked features and the hydrological feature map.

[0040] The second fully connected layer is used to obtain the prediction results of hydrological time series.

[0041] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the hydrological time series prediction method described in the first aspect.

[0042] Fourthly, the present invention provides a computer device, comprising:

[0043] Memory, used to store computer instructions;

[0044] A processor for executing the computer instructions to implement the steps of the hydrological time series prediction method described in the first aspect.

[0045] Beneficial effects

[0046] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0047] This invention expands the receptive field through temporal convolution, thereby broadening the feature extraction range. The extracted features can cover information from multiple effective time points. By encoding relative position information through a rotation matrix, the original data vector is directly rotated to mine the dependencies between adjacent data, making full use of the original data. Without adding position vectors, the prediction effect is good. Attached Figure Description

[0048] Figure 1 The diagram shown is a flowchart of one embodiment of the hydrological time series prediction method of the present invention.

[0049] Figure 2 The diagram shown is a schematic representation of the hydrological time series prediction system of the present invention in one embodiment.

[0050] Figure 3 The diagram shown is a structural schematic of the temporal convolution module of the present invention in one embodiment;

[0051] Figure 4 The diagram shown is a comparison of the prediction results of different hydrological prediction models in one embodiment of the present invention.

[0052] Figure 5 The diagram shows a comparison of loss curves for different hydrological prediction models in one embodiment of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0054] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0055] Example 1

[0056] like Figure 1 As shown in the figure, this embodiment introduces a method for predicting hydrological time series data, including:

[0057] Obtain hydrological time series data. In this embodiment, continuous data over a certain period of time is obtained from a pumping station.

[0058] Missing values ​​in the hydrological sequence were filled using the near-distance interpolation method.

[0059] When the missing value of the hydrological sequence is a single value, the average of the two values ​​before and after the missing value is taken as the missing value.

[0060] When the missing values ​​of the hydrological sequence are multi-valued, the formula for the missing values ​​is:

[0061] ;

[0062] In the formula, For the first One missing value. ; The value preceding the missing value; The value following the missing value; The number of missing values; This represents the number of values ​​preceding the missing value.

[0063] The hydrological time series is sequentially subjected to temporal convolution and dimensional adjustment to obtain the dimensionally adjusted hydrological features.

[0064] Hydrological features are obtained by performing temporal convolution on the hydrological time series using a temporal convolution module. In this embodiment, the temporal convolution module can adopt the TCN (Temporal Convolutional Networks) model.

[0065] The temporal convolution module includes multiple cascaded residual blocks, and the dilation rate of these cascaded residual blocks increases sequentially. For example... Figure 3 The diagram shows the structure of the temporal convolution module. In this embodiment, there are three residual blocks, and their dilation rates d are 1, 2, and 4 respectively.

[0066] Each residual block includes a dilated causal convolutional layer, a weight normalization layer, a ReLU activation function layer, a Dropout function layer, and a residual structure layer. Different residual blocks have different dilation rates; the larger the dilation rate, the larger the receptive field and the more features are extracted.

[0067] The execution steps for each residual block are as follows:

[0068] Hydrological time series data are processed through dilated causal convolutional layers, where the convolutional kernels perform cross-connections to expand the receptive field, thereby broadening the scope of feature extraction.

[0069] Then, the network is processed by a weight normalization layer to eliminate dimensionality and reduce internal covariate bias, thereby accelerating the convergence speed of the network.

[0070] After passing through the ReLU activation function layer and the Dropout function layer, the ReLU activation function adds non-linear properties to the residual block, and the Dropout function randomly suppresses certain features, enhancing the generalization ability of the residual block.

[0071] Residual connections are introduced into each residual block, and the formula for calculating the hydrological characteristics is as follows:

[0072] ;

[0073] In the formula, Hydrological characteristics; For residual mapping; This is a hydrological time series; It is the first weight; It is the second weight.

[0074] The hydrological time series is dimensionally adjusted by the first fully connected layer to obtain the dimensionally adjusted hydrological features.

[0075] The dimensionally adjusted hydrological features are input into a pre-trained hydrological prediction model, which outputs the predicted hydrological time series data. The difference between the predicted result in each round and the actual data in the training set is calculated using the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) loss functions. Knowing the difference, the model parameters are adjusted by an optimizer. After multiple rounds of adjustment, the predicted result is made closest to the true value. The hydrological prediction model can be a Trpe-Informer model, which is an Informer model encoded with relative positions using a rotation matrix.

[0076] The processing steps of the hydrological prediction model include:

[0077] The relative position of the dimension-adjusted hydrological features is encoded using a rotation matrix to obtain hydrological features containing location information.

[0078] In two-dimensional space, a rotation matrix is ​​obtained by rotating a vector by an angle to obtain a new vector. The formula for calculating the rotation matrix is:

[0079] ;

[0080] In the formula, The angle is the rotation angle.

[0081] The calculation formula for the hydrological features containing location information is as follows:

[0082] ;

[0083] In the formula, For the first A hydrological feature matrix containing location information; For the first Hydrological feature matrix adjusted in each dimension; For matrix The first component; For matrix The second component; For the first Rotation matrix at each position; For the first The rotation angle at each position.

[0084] When performing sparse correlation calculations, the correlation result between two tokens depends on nm, meaning it is influenced by their relative positions, which can improve the utilization of data relationships. For example, the formula is:

[0085] ;

[0086] In the formula, For the first A hydrological feature matrix containing location information; For the first A rotation matrix for each position.

[0087] The hydrological features containing location information can be used to determine which data have a strong correlation, or which data are most relevant to the data to be predicted, thus facilitating subsequent sparse correlation calculations.

[0088] An encoder is used to calculate sparse correlation and compress features from hydrological features containing location information, resulting in a hydrological feature map. Specifically:

[0089] The hydrological features containing location information are added to the sparse attention mechanism of the encoder, and sparse correlation calculation is performed on them. The final result is features containing location information, which makes full use of the correlation of different features.

[0090] Features containing location information are compressed through a distillation process to obtain a hydrological feature map.

[0091] The encoder improves its robustness by stacking the two steps mentioned above.

[0092] A decoder is used to perform masking operations and sparse correlation calculations on hydrological features containing location information to obtain masked hydrological features. The masked hydrological features are then compared with the hydrological feature map to obtain an output feature map.

[0093] Masking operations are used to conceal certain feature values, preventing them from affecting parameter updates. For a sequence, our decoded output can only depend on the output before time t, not on the output after t. Therefore, we need to use masking operations to hide the information after t.

[0094] The sparse correlation between the masked hydrological features and the hydrological feature map is calculated to obtain the output feature map.

[0095] The formula for calculating the sparse correlation is:

[0096] ;

[0097] In the formula, For query matrix; The key matrix; It is the eigenvalue matrix; For the Softmax function; for and KL divergence analysis was used to determine the top activity levels. indivual ; It is the transpose of the key matrix; The dimension of the feature vector.

[0098] The output feature map is predicted using a second fully connected layer to obtain the prediction results of the hydrological time series.

[0099] Example 2

[0100] like Figure 2 As shown in the figure, this embodiment introduces a hydrological time series prediction system, including:

[0101] The data acquisition module is used to acquire hydrological time series data.

[0102] The temporal convolution module is used to perform temporal convolution on hydrological time series.

[0103] The first fully connected layer is used to adjust the dimensions of the hydrological time series.

[0104] A hydrological prediction model is used to obtain prediction results for hydrological time series. The hydrological prediction model includes an encoder, a decoder, and a second fully connected layer.

[0105] The encoder is used to perform sparse correlation calculation and feature compression on hydrological features containing location information.

[0106] The decoder is used to perform sparse correlation calculation on the masked hydrological features containing location information, and to perform sparse correlation calculation on the masked features and the hydrological feature map.

[0107] The second fully connected layer is used to obtain the prediction results of hydrological time series.

[0108] Example 3

[0109] like Figure 4 , 5 As shown in the figure, this embodiment provides an experimental design for a hydrological time series prediction method, which specifically includes the following steps:

[0110] Continuous data over a certain period of time is obtained from a certain pumping station to form a hydrological time series. The unit power, pumping station unit speed, flow rate, outlet pressure, main bearing temperature, and upstream water level data of the Jinhu Station of the South-to-North Water Diversion East Route are selected to predict the water level of the pumping station.

[0111] First, use the near-interpolation method for preprocessing to ensure that there is data at each time point.

[0112] This embodiment uses minute-by-minute data from May 6th to May 9th, 2017, covering the unit power, pump station speed, flow rate, outlet pressure, main bearing temperature, and upstream water level at the Huai'an Jinhu Pumping Station. A total of 10,000 data points are used to train the model and predict the upstream water level. Normalization is performed using the min-max normalization method, as shown in the following formula:

[0113] ;

[0114] In the formula, This is the data after normalization. This is the original data; The maximum value in the original data; This is the minimum value in the original data.

[0115] The training data from Jinhu Station is input into the Temporal Convolutional Network (TCN) module, using three residual blocks concatenated, such as... Figure 3 As shown, x represents the input, y represents the output, and the dilation rates are d=1, 2, and 4. By expanding the receptive field through multi-layer convolution, the feature extraction range is broadened, allowing the final data to cover information from multiple effective time points. The execution steps for each residual block are as follows:

[0116] After receiving data, the residual block undergoes dilated causal convolution. The convolution kernel performs cross-connection to increase the receptive field, expand the feature extraction range, and make full use of historical data from different times.

[0117] Weight normalization eliminates dimensionality and reduces internal covariate bias, thereby accelerating network convergence.

[0118] The ReLU activation function and the Dropout function are used. The activation function adds non-linear properties to neurons, while the Dropout function randomly suppresses certain features, enhancing the model's generalization ability.

[0119] Introducing residual connectivity, the formula for calculating the hydrological characteristics is as follows:

[0120] ;

[0121] In the formula, Hydrological characteristics; For residual mapping; This is a hydrological time series; It is the first weight; It is the second weight.

[0122] The hydrological features are fed into a fully connected layer, and the output dimension is adjusted to obtain the dimension-adjusted hydrological features.

[0123] The dimensionally adjusted hydrological features are input into a pre-trained hydrological prediction model, which outputs the prediction results for the hydrological time series. The hydrological prediction model is continuously trained and optimized through backpropagation. In this embodiment, the hydrological prediction model can be a Trpe-Informer model.

[0124] In this embodiment, the Trpe-Informer model uses the Adam optimizer with an initial learning rate of 0.002. The learning rate is halved in each epoch. The expansion rate d of the TCN model is 1, 2, and 4 respectively. The input size of both the encoder and decoder is set to 6, which takes six key features as input: receiver power, pump station unit speed, flow rate, outlet pressure, main bearing temperature, and upstream water level of the pump station. The output size of the encoder and decoder is 1, which is the output of the upstream water level of the pump station.

[0125] To verify the predictive effectiveness of this invention, data from the Jinhu Pumping Station were selected and compared with the LSTM, Transformer, and Informer models using the hydrological prediction model of this invention, namely the Trpe-Informer model. The evaluation metrics for the model prediction results were the MAE loss function and the RMSE loss function. The results are as follows:

[0126] Table 1 Performance Comparison of Different Models

[0127]

[0128] A comparison of the data in Table 1 clearly shows that the hydrological prediction model of this invention has relatively smaller errors compared to the other three models. Overall, the prediction accuracy of this invention is higher. Figure 4 The image shows a comparison of the prediction results of different models.

[0129] like Figure 5 As shown, by comparing the Informer and Trpe-Informer models, the loss value gradually stabilizes as the number of training iterations increases during training. Moreover, the loss curve of the Trpe-Informer model converges faster because the Trpe-Informer model uses a rotation matrix for relative position encoding. Each position in the sequence is represented by a rotation in the embedding space, which is faster than sinusoidal embedding.

[0130] Example 4

[0131] This embodiment introduces a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the hydrological time series prediction method described in Embodiment 1 or 3.

[0132] Example 5

[0133] This embodiment describes a computer device, including:

[0134] Memory, used to store computer instructions;

[0135] A processor for executing the computer instructions to implement the steps of the hydrological time series prediction method described in Embodiment 1 or 3.

[0136] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0138] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0140] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for predicting hydrological time series data, characterized in that, include: Obtain hydrological time series; The hydrological time series is sequentially subjected to temporal convolution and dimensionality adjustment to obtain the dimensionally adjusted hydrological features. The dimensionally adjusted hydrological features are input into a pre-trained hydrological prediction model, which outputs the prediction results of the hydrological time series. The hydrological prediction model adopts the Trpe-Informer model, which is obtained by encoding the relative position of the Informer model through a rotation matrix. The processing steps of the hydrological prediction model include: The relative position of the dimension-adjusted hydrological features is encoded using a rotation matrix to obtain hydrological features containing location information. An encoder is used to perform sparse correlation calculation and feature compression on hydrological features containing location information to obtain a hydrological feature map. A decoder is used to perform masking operations and sparse correlation calculations on hydrological features containing location information to obtain masked hydrological features. The masked hydrological features are then compared with the hydrological feature map to obtain an output feature map. The output feature map is predicted using a second fully connected layer to obtain the prediction results of the hydrological time series. Before performing temporal convolution on the hydrological time series, missing values ​​in the hydrological time series are filled using the near-distance interpolation method. When the missing value of the hydrological time series is a single value, the average of the two values ​​before and after the missing value is taken as the missing value; When the missing values ​​of the hydrological time series are multi-valued, the formula for the missing values ​​is: ； In the formula, For the first One missing value. ; The value preceding the missing value; The value following the missing value; The number of missing values; This represents the number of values ​​preceding the missing value. Hydrological features are obtained by performing temporal convolution on the hydrological time series using a temporal convolution module. The temporal convolution module includes multiple interconnected residual blocks, and the dilation rate of the multiple interconnected residual blocks increases sequentially. Residual connections are introduced into each residual block, and the formula for calculating the hydrological characteristics is as follows: ； In the formula, Hydrological characteristics; For residual mapping; This is a hydrological time series; It is the first weight; As the second weight; The formula for calculating the rotation matrix is: ； In the formula, The rotation angle; The calculation formula for the hydrological features containing location information is as follows: ； In the formula, For the first A hydrological feature matrix containing location information; For the first Hydrological feature matrix adjusted in each dimension; For matrix The first component; For matrix The second component; For the first Rotation matrix at each position; For the first The rotation angle at each position.

2. The method for predicting hydrological time series according to claim 1, characterized in that, The formula for calculating the sparse correlation is: ； In the formula, For query matrix; The key matrix; It is the eigenvalue matrix; For the Softmax function; for and KL divergence analysis was used to determine the top activity levels. indivual ; It is the transpose of the key matrix; The dimension of the feature vector.

3. A hydrological time series prediction system, characterized in that, The system for performing the hydrological time series prediction method according to any one of claims 1 to 2 includes: The data acquisition module is used to acquire hydrological time series data. The temporal convolution module is used to perform temporal convolution on hydrological time series. The first fully connected layer is used for dimensional adjustment of hydrological time series. The encoder is used to perform sparse correlation calculation and feature compression on hydrological features containing location information; The decoder is used to perform sparse correlation calculation on the masked hydrological features containing location information, and to perform sparse correlation calculation on the masked features and the hydrological feature map. The second fully connected layer is used to obtain the prediction results of hydrological time series.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the hydrological time series prediction method according to any one of claims 1-2.

5. A computer device, characterized in that, include: Memory, used to store computer instructions; A processor for executing the computer instructions to implement the steps of the hydrological time series prediction method according to any one of claims 1-2.

Images