Runoff prediction method based on convolution long-short term memory network and wavelet fusion

Abstract

The invention relates to a runoff prediction method based on convolution long and short time memory network and wavelet fusion. The runoff prediction method comprises the following steps: (1) data preprocessing; (2) extracting spatio-temporal characteristics; (3) extracting time sequence features; (4) feature compression and alignment; (5) carrying out cross-modal feature fusion; and (6) runoff prediction. According to the method, collaborative modeling of a meteorological element space interaction relationship and runoff multi-scale time sequence characteristics can be realized, the defect of spatial-temporal characteristic splitting processing of a traditional model is overcome, the responsiveness of the model to runoff changes under complex meteorological conditions is enhanced through a dynamic characteristic fusion mechanism, and the generalization and accuracy of a prediction model are improved.

Description

Technical Field

[0001] This invention relates to a runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet, which is applicable to smart water conservancy, flood control scheduling and water resource management, and belongs to the field of watershed runoff prediction. Background Technology

[0002] Watershed runoff forecasting is of great significance for water resource management, flood control and disaster reduction, and water conservancy project planning. Accurate runoff forecasting can provide a scientific basis for reservoir scheduling and irrigation water allocation, thereby optimizing water resource utilization efficiency and reducing losses caused by floods and droughts.

[0003] Traditional runoff forecasting methods primarily rely on empirical statistical models or hydrological models based on physical mechanisms. Empirical statistical models, such as the unit hydrograph method and the SCS curve number method, while simple and easy to use, struggle to capture the impact of complex meteorological dynamics on runoff, resulting in lower prediction accuracy under atypical or extreme weather conditions. Hydrological models based on physical mechanisms, such as the SWAT model and the VIC model, while capable of simulating watershed hydrological cycles well, require a large number of watershed underlying surface parameters that are difficult to obtain accurately. Furthermore, parameter calibration is typically very complex and computationally intensive, making it particularly difficult to meet the real-time forecasting requirements of scenarios demanding rapid response, such as real-time flood forecasting. In recent years, with the development of deep learning technology, neural network-based runoff forecasting models have been widely applied, achieving significant progress in accuracy improvement, adaptability to complex scenarios, and the integration of multiple technologies.

[0004] However, existing neural network-based runoff prediction methods often treat the spatiotemporal characteristics of meteorological elements and the temporal characteristics of runoff sequences independently, neglecting the intrinsic connection between the spatial interactions of meteorological elements and the multi-scale temporal characteristics of runoff. For example, some methods simply input observational data from meteorological stations into the neural network, ignoring the impact of the spatial distribution and dynamic evolution of meteorological elements within the watershed on runoff. Other methods, while considering the temporal characteristics of runoff sequences, fail to effectively extract the multi-timescale information contained within the runoff sequences, such as long-term trends, seasonal fluctuations, and short-term fluctuations. This deficiency in the fragmented treatment of spatiotemporal characteristics limits the model's ability to respond to runoff changes under complex meteorological conditions, reducing the robustness and accuracy of runoff prediction. Therefore, it is necessary to propose a runoff prediction method that simultaneously considers the spatiotemporal characteristics of meteorological elements and the multi-scale temporal characteristics of runoff sequences. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a runoff prediction method based on the fusion of convolutional long short-term memory networks and wavelets. By combining data processing and analysis techniques and fully considering the dynamic correlation of data, the accuracy of runoff prediction is improved.

[0006] This invention provides a technical solution: a runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet, comprising the following steps: (1) Data preprocessing: The collected watershed runoff data and meteorological data from various meteorological stations are standardized and gridded to obtain standardized daily runoff sequence data and gridded meteorological data; (2) Extracting spatiotemporal features: Using a convolutional long short-term memory network, the gridded meteorological data is converted into gridded spatiotemporal features, and the dynamic evolution law of meteorological elements is extracted to obtain the spatiotemporal feature tensor of meteorological data; (3) Extracting temporal features: Wavelet transform is used to decompose the standardized daily runoff sequence data into multiple layers to generate feature components at multiple time scales; (4) Feature compression and alignment: The spatiotemporal feature tensor of meteorological data is compressed by global pooling operation to obtain the compressed spatiotemporal feature sequence of meteorological data. The time step of wavelet decomposition coefficients is made consistent with the time step of the spatiotemporal feature tensor of meteorological data by the time alignment module to obtain the aligned wavelet feature sequence. (5) Cross-modal feature fusion: The spatiotemporal feature sequence of compressed meteorological data and the aligned wavelet feature sequence are fused through the cross-modal attention mechanism to form a fused feature sequence that integrates historical meteorological data and runoff. (6) Runoff prediction: Runoff prediction is performed by learning the dynamic mapping relationship between the fusion feature sequence and runoff through a long short-term memory network.

[0007] Further, in step (1), the collected watershed runoff data and meteorological data of each meteorological station are standardized and gridded. The standardization process includes: identifying outliers in the watershed daily runoff data and daily meteorological data of each meteorological station in the watershed using statistical or visualization methods, filling out missing values ​​or using modeling methods, correcting, replacing or deleting outliers, and performing consistency processing on the data to obtain standardized daily runoff sequence data and standardized meteorological data.

[0008] Further, in step (1), the collected watershed runoff data and meteorological data from various meteorological stations are standardized and gridded. The gridding includes: collecting watershed boundary vector data, dividing the watershed into 0.1°×0.1° grids by combining latitude and longitude coordinates, and using an inverse distance weighted interpolation algorithm to interpolate the standardized meteorological data to the grid cells to generate a gridded meteorological data tensor, i.e., gridded meteorological data.

[0009] Further, in step (2), the gridded meteorological data is converted into gridded spatiotemporal features using a convolutional long short-term memory network, and the dynamic evolution law of meteorological elements is extracted to obtain the spatiotemporal feature tensor of meteorological data. This includes: constructing a ConvLSTM model, the input of which is gridded meteorological data, and the output is the spatiotemporal feature tensor of meteorological data. The spatiotemporal feature tensor of meteorological data includes the dynamic correlation information of meteorological elements in time and space.

[0010] Further, in step (3), wavelet transform is used to decompose the runoff sequence data into multiple layers to generate feature components of multiple time scales, including: using the db4 wavelet basis function and Mallat algorithm to decompose the standardized daily runoff sequence data into three layers to obtain the detailed components D1 of daily-weekly scale fluctuations, the detailed components D2 of weekly-monthly scale fluctuations, the detailed components D3 of monthly-seasonal scale fluctuations, and the approximate component A3 of annual scale trend.

[0011] Further, in step (4), the spatiotemporal feature tensor of meteorological data is compressed by global pooling operation to obtain a compressed spatiotemporal feature sequence of meteorological data, including: performing global average pooling on the spatiotemporal feature tensor of meteorological data to compress the spatial dimension to 1×1, thereby obtaining a compressed spatiotemporal feature sequence of meteorological data.

[0012] Further, in step (4), the time step of the wavelet decomposition coefficients is aligned with the time step of the spatiotemporal feature tensor of the meteorological data through the time alignment module to obtain the aligned wavelet feature sequence, including: each component after wavelet decomposition is zero-padded and then the wavelet component is truncated by a sliding window so that each window contains wavelet components for n consecutive days. The coefficient distribution of the edge window is adjusted by linear interpolation to ensure that the time step of each wavelet component window is aligned with the time step of the spatiotemporal feature tensor of the meteorological data, thereby generating the aligned wavelet feature sequence.

[0013] Further, in step (5), the spatiotemporal feature sequence of compressed meteorological data and the aligned wavelet feature sequence are fused through a cross-modal attention mechanism to form a fused feature sequence that integrates historical meteorology and runoff. This includes: mapping the aligned wavelet feature sequence to the same dimension as the spatiotemporal feature sequence of compressed meteorological data through a linear transformation, and calculating the mapped wavelet feature sequence and the spatiotemporal feature sequence of compressed meteorological data using the scaling dot product attention formula. The resulting output matrix is ​​added to the spatiotemporal feature sequence of compressed meteorological data to obtain the fused feature sequence.

[0014] Due to the application of the above technical solution, the present invention has the following advantages compared with the prior art: The present invention provides a runoff prediction method based on convolutional long short-term memory network and wavelet fusion, which can realize the collaborative modeling of spatial interaction relationship of meteorological elements and multi-scale temporal characteristics of runoff. It overcomes the defects of traditional models in processing spatiotemporal features separately. Through dynamic feature fusion mechanism, it enhances the model's response to runoff changes under complex meteorological conditions and improves the generalization and accuracy of prediction model. Detailed Implementation

[0015] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0017] This embodiment presents a runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet, including the following steps: (1) Data preprocessing: The collected watershed runoff data and meteorological data from various meteorological stations were standardized and gridded to obtain standardized daily runoff sequence data and gridded meteorological data. Specifically, this included: ① Standardization processing: For daily runoff data (collected from the hydrological station at the outlet of the basin) and daily meteorological data (including daily precipitation, temperature, humidity, sunshine duration, etc.) from various meteorological stations within the basin, statistical or visualization methods are used to identify outliers. Missing values ​​are filled or modeled, and outliers are corrected, replaced, or deleted. The data is then processed for consistency.

[0018] The watershed daily runoff data were standardized using Z-score (|Z|>3 is considered anomaly): , Where μ is the mean and σ is the standard deviation, the data is mapped to a distribution with a mean of 0 and a standard deviation of 1 to eliminate the influence of dimensions and remove outliers exceeding 3 times the standard deviation to obtain standardized daily runoff sequence data.

[0019] Min-max normalization was applied to the daily meteorological data from each meteorological station within the basin. , The data is compressed to the [0,1] interval, preserving the relative distribution characteristics of the original data, to obtain standardized meteorological data.

[0020] ② Mesh processing: Watershed boundary vector data were collected and, combined with latitude and longitude coordinates, the watershed was divided into a 0.1° × 0.1° grid (approximately 11km × 11km, with slight variations in size depending on latitude), covering the entire watershed area. The inverse distance weighted (IDW) interpolation algorithm was used to interpolate the standardized meteorological data obtained in step ① above to the grid cells, generating a gridded meteorological data tensor (i.e., gridded meteorological data). The dimensions of the gridded meteorological data tensor are: time step × number of grid rows × number of grid columns × number of meteorological elements.

[0021] (2) Extracting spatiotemporal features: Using convolutional long short-term memory network, gridded meteorological data is converted into gridded spatiotemporal features, and the dynamic evolution law of meteorological elements is extracted to obtain the spatiotemporal feature tensor of meteorological data.

[0022] Specifically, this includes: constructing a ConvLSTM model, which consists of an input layer, a first ConvLSTM layer, a second ConvLSTM layer, and an output layer set up sequentially.

[0023] Input layer: Receives gridded meteorological data, i.e., the gridded meteorological data tensor obtained in step ② above (a sliding window with a time step of n days, i.e., each sample contains gridded meteorological data for n consecutive days, with dimensions of: time step n × number of grid rows a × number of grid columns b × number of meteorological elements c).

[0024] The first ConvLSTM layer: sets d 3×3 convolution kernels with a stride of 1 and padding mode of "same". The output dimension is n×a×b×d, capturing the spatiotemporal correlation of local meteorological elements within the grid.

[0025] The second ConvLSTM layer: sets 2d 3×3 convolutional kernels with a stride of 1 and padding mode of "same". The output dimension is n×a×b×2d, which extracts the dynamic evolution of meteorological patterns over a larger range.

[0026] Output layer: Outputs the spatiotemporal feature tensor of meteorological data, i.e., gridded spatiotemporal features with dimensions of n×a×b×2d, which contains the dynamic correlation information of meteorological elements in time and space.

[0027] (3) Extracting time-series features: Wavelet transform is used to decompose the standardized daily runoff sequence data into multiple layers to generate feature components at multiple time scales.

[0028] Specifically, this includes: using the db4 wavelet basis function, applying the Mallat algorithm to perform a 3-level decomposition of standardized diurnal runoff series data. The decomposition process (Mallat algorithm flow) is as follows: Layer 1: The original sequence S0 is convolved and downsampled by a low-pass filter h and a high-pass filter g to obtain the low-frequency approximate component A1 (reflecting the medium- to long-term trend) and the high-frequency detail component D1 (reflecting short-term fluctuations). Layer 2: Repeat the above operation on A1 to obtain A2 (smoother trend) and D2 (medium-term fluctuation). Level 3: Decompose A2 to obtain A3 (long-term trend) and D3 (seasonal fluctuation).

[0029] Finally, the detailed components D1 of the daily-weekly scale fluctuations, the detailed components D2 of the weekly-monthly scale fluctuations, the detailed components D3 of the monthly-seasonal scale fluctuations, and the approximate component A3 of the annual scale trend are obtained.

[0030] (4) Feature Compression and Alignment: The spatiotemporal feature tensor of meteorological data is compressed through global pooling to obtain a compressed spatiotemporal feature sequence of meteorological data. Then, the time step of the wavelet decomposition coefficients is aligned with the time step of the spatiotemporal feature tensor of the meteorological data using a time alignment module to obtain an aligned wavelet feature sequence. The specific method is as follows: Global pooling compression: Global average pooling is performed on the spatiotemporal feature tensor (n×a×b×2d) of meteorological data output by the ConvLSTM model to compress the spatial dimension to 1×1, resulting in the compressed spatiotemporal feature sequence, i.e., the compressed meteorological data spatiotemporal feature sequence.

[0031] .

[0032] Time alignment: After zero-padding, each component of the wavelet decomposition is truncated using a sliding window so that each window contains wavelet components from n consecutive days. The coefficient distribution of the edge windows is adjusted by linear interpolation to ensure that the time step of each wavelet component window is consistent with the time step (i.e., n) of the spatiotemporal feature tensor of the meteorological data, thus generating an aligned wavelet feature sequence.

[0033] .

[0034] (5) Cross-modal feature fusion: The spatiotemporal feature sequence of compressed meteorological data and the aligned wavelet feature sequence are fused through a cross-modal attention mechanism to form a fused feature sequence that integrates historical meteorological data and runoff data. The specific method is as follows: The two input sequences are the compressed meteorological data spatiotemporal feature sequence (dimension n×2d) and the aligned wavelet feature sequence (dimension n×4). Since the two sequences have different dimensions, the aligned wavelet feature sequence is first mapped to the same dimension as the compressed meteorological data spatiotemporal feature sequence through a linear transformation. The mapped wavelet feature sequence and the compressed meteorological data spatiotemporal feature sequence are then calculated using the scaling dot product attention formula: , Where Q is the spatiotemporal feature sequence matrix of compressed meteorological data (n×2d), K is the mapped wavelet feature sequence matrix (n×2d), and V is the value matrix of the mapped wavelet feature sequence (n×2d). k =2d represents the dimensions of Q and K, and T represents the matrix transpose operation.

[0035] Attention output matrix Adding this to the meteorological spatiotemporal feature sequence (here referring to the meteorological data spatiotemporal feature tensor after global pooling compression) yields the final fused feature sequence: , The fusion feature sequence (n×2d) simultaneously includes meteorological spatiotemporal dynamics and runoff multi-scale temporal features.

[0036] (6) Runoff Prediction: Runoff prediction is performed by learning the dynamic mapping relationship between the fused feature sequences and runoff through a Long Short-Term Memory (LSTM) network. The specific method is as follows: An input layer, a first LSTM layer, a second LSTM layer, and an output layer are set up, where: Input layer: Receives the fused feature sequence (n×2d) obtained in step (5); First LSTM layer: Set the number of hidden units to be twice the number of the second ConvLSTM layer convolution kernel in step (2), 4d, and return the sequence (output n×4d). Second LSTM layer: Set the same number of hidden units as the second ConvLSTM layer in step (2), 2d, and return the output of the last time step (1×2d). Output layer: Fully connected layer (with N neurons), which outputs the predicted runoff values ​​for the next N days through a linear activation function (the actual flow values ​​are obtained after inverse standardization).

[0037] Therefore, due to the application of the above technical solution, the present invention has the following advantages compared with the prior art: This paper presents a runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet, which can realize the collaborative modeling of the spatial interaction relationship of meteorological elements and the multi-scale temporal characteristics of runoff. It overcomes the shortcomings of traditional models in processing spatiotemporal features separately. Through the dynamic feature fusion mechanism, it enhances the model's response to runoff changes under complex meteorological conditions and improves the generalization and accuracy of the prediction model.

[0038] The above embodiments are only for illustrating the technical concept and features of the present invention. Their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be used to limit the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet, characterized in that: Includes the following steps: (1) Data preprocessing: The collected watershed runoff data and meteorological data from various meteorological stations are standardized and gridded to obtain standardized daily runoff sequence data and gridded meteorological data; (2) Extracting spatiotemporal features: The gridded meteorological data is converted into gridded spatiotemporal features using a convolutional long short-term memory network, and the dynamic evolution law of meteorological elements is extracted to obtain the spatiotemporal feature tensor of meteorological data; (3) Extracting temporal features: Wavelet transform is used to decompose the standardized daily runoff sequence data into multiple layers to generate feature components at multiple time scales; (4) Feature compression and alignment: The spatiotemporal feature tensor of the meteorological data is compressed by global pooling operation to obtain a compressed spatiotemporal feature sequence of meteorological data. The time step of the wavelet decomposition coefficients is made consistent with the time step of the spatiotemporal feature tensor of the meteorological data by the time alignment module to obtain an aligned wavelet feature sequence. (5) Cross-modal feature fusion: The spatiotemporal feature sequence of the compressed meteorological data and the aligned wavelet feature sequence are fused through a cross-modal attention mechanism to form a fused feature sequence that integrates historical meteorology and runoff; (6) Runoff prediction: Runoff prediction is performed by learning the dynamic mapping relationship between the fused feature sequence and runoff through a long short-term memory network.

2. The runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet as described in claim 1, characterized in that: In step (1), the collected watershed runoff data and meteorological data from various meteorological stations are standardized and gridded. The standardization process includes: identifying outliers in the watershed daily runoff data and daily meteorological data from various meteorological stations within the watershed using statistical or visualization methods, filling out missing values ​​using imputation or modeling methods, correcting, replacing, or deleting outliers, and performing consistency processing on the data to obtain standardized daily runoff sequence data and standardized meteorological data.

3. The runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet as described in claim 2, characterized in that: In step (1), the collected watershed runoff data and meteorological data from various meteorological stations are standardized and gridded. The gridding includes: collecting watershed boundary vector data, dividing the watershed into 0.1°×0.1° grids by combining latitude and longitude coordinates, and using an inverse distance weighted interpolation algorithm to interpolate the standardized meteorological data to the grid cells to generate a gridded meteorological data tensor, i.e., the gridded meteorological data.

4. The runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet as described in claim 1, characterized in that: In step (2), the process of using a convolutional long short-term memory network to convert the gridded meteorological data into gridded spatiotemporal features, extracting the dynamic evolution law of meteorological elements, and obtaining a meteorological data spatiotemporal feature tensor includes: constructing a ConvLSTM model, the input of which is the gridded meteorological data, and the output is a meteorological data spatiotemporal feature tensor, wherein the meteorological data spatiotemporal feature tensor includes the dynamic correlation information of meteorological elements in time and space.

5. The runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet as described in claim 1, characterized in that: In step (3), the use of wavelet transform to perform multi-level decomposition of runoff sequence data to generate feature components of multiple time scales includes: using the db4 wavelet basis function and the Mallat algorithm to perform 3-level decomposition of the standardized daily runoff sequence data to obtain the detailed components D1 of daily-weekly scale fluctuations, the detailed components D2 of weekly-monthly scale fluctuations, the detailed components D3 of monthly-seasonal scale fluctuations, and the approximate component A3 of annual scale trend.

6. A runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet as described in claim 1 or 4, characterized in that: In step (4), the compression of the spatiotemporal feature tensor of the meteorological data through global pooling operation to obtain the compressed spatiotemporal feature sequence of meteorological data includes: performing global average pooling on the spatiotemporal feature tensor of the meteorological data to compress the spatial dimension to 1×1, thereby obtaining the compressed spatiotemporal feature sequence of meteorological data.

7. A runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet as described in any one of claims 1, 4, or 5, characterized in that: In step (4), the time alignment module is used to make the time step of the wavelet decomposition coefficients consistent with the time step of the spatiotemporal feature tensor of the meteorological data to obtain the aligned wavelet feature sequence. This includes: after zero-padding, each component after wavelet decomposition is truncated by a sliding window so that each window contains wavelet components for n consecutive days. The coefficient distribution of the edge window is adjusted by linear interpolation to ensure that the time step of each wavelet component window is consistent with the time step of the spatiotemporal feature tensor of the meteorological data, thereby generating the aligned wavelet feature sequence.

8. The runoff prediction method based on the fusion of convolutional long short-term memory network and wavelet as described in claim 1, characterized in that: In step (5), the process of fusing the spatiotemporal feature sequence of the compressed meteorological data and the aligned wavelet feature sequence through a cross-modal attention mechanism to form a fused feature sequence that integrates historical meteorology and runoff includes: mapping the aligned wavelet feature sequence to the same dimension as the spatiotemporal feature sequence of the compressed meteorological data through a linear transformation, and calculating the mapped wavelet feature sequence and the spatiotemporal feature sequence of the compressed meteorological data using a scaling dot product attention formula, and adding the resulting output matrix to the spatiotemporal feature sequence of the compressed meteorological data to obtain the fused feature sequence.