Hydrological time series prediction network and method based on feature extraction and guidance

Abstract

The invention provides a hydrological time series prediction network and method based on feature extraction and guidance, and belongs to the technical field of hydrological time prediction. The problems of insufficient prediction precision, fast error speed increase and the like of the existing method are solved; an encoder-decoder model structure of Informer is adopted, the system further comprises a feature extraction module, and the output of the feature extraction module serves as the input of the encoder; the feature extraction module comprises a time domain feature extraction module, a frequency domain feature extraction module and a dynamic feature fusion module, and the time domain feature extraction module is used for extracting multi-scale time domain features in the hydrological time sequence data; the frequency domain feature extraction module extracts global frequency domain features in the hydrological time sequence data through discrete cosine transform; the dynamic feature fusion module is used for performing channel-level and space-level fusion on the time domain features and the frequency domain features; an attention guiding mechanism is adopted to replace a sparse attention mechanism in an original encoder; the method is applied to hydrological time series prediction.

Description

Technical Field

[0001] The present application relates to the technical field of hydrological time prediction, and particularly relates to a hydrological time series prediction network and method based on feature extraction and guidance. Background Art

[0002] Hydrological time series prediction is a core research direction in the field of hydrometeorology and is widely applied to key scenarios such as watershed runoff management, flood control and disaster reduction, and water resources scheduling. Accurately predicting the changing trends of runoff and meteorological variables (such as precipitation, temperature, humidity, etc.) is of great significance for resource management and scientific decision-making. However, hydrometeorological data usually has non-linear, non-stationary, and multi-scale dynamic characteristics, which significantly increase the complexity of time series prediction, and it is difficult for traditional methods to effectively handle.

[0003] Traditional time series prediction methods mainly include the ETS and ARIMA methods based on statistical models, as well as support vector regression and decision tree models based on machine learning. Statistical methods have certain advantages in dealing with short-term and linear sequences, but it is difficult to handle complex data with non-linearity and multiple variables; although machine learning methods can capture non-linear relationships, it is difficult to characterize the dynamic dependencies in time series, especially showing obvious limitations in long time series tasks.

[0004] In recent years, deep learning technology has gradually become the mainstream means for time series prediction. Methods such as recursive neural network (RNN), long short-term memory network (LSTM), and gated recurrent unit (GRU) provide diverse choices for time series modeling, but problems such as gradient disappearance and long training time are likely to occur when dealing with long sequence data. The Transformer model overcomes these limitations through a global self-attention mechanism, can capture long-term dependencies and achieve parallel computing, significantly improving the prediction efficiency. However, its computational complexity grows with the square of the sequence length, limiting its practical application in long time series prediction. To address this problem, the Informer model proposes a sparse attention mechanism, which significantly reduces the computational complexity by selectively retaining key feature points and shows advantages in long sequence prediction.

[0005] Although the Informer model has improved in terms of efficiency and modeling ability, it is still difficult to fully capture complex dynamic changes, such as periodic patterns, non-stationary changes, and multi-variable interaction features in hydrometeorological data. In addition, single time domain or frequency domain modeling methods are difficult to mine the multi-scale features of data, resulting in insufficient prediction accuracy and stability of the model in complex scenarios. Summary of the Invention

[0006] In order to solve the problems of insufficient prediction accuracy and rapid error growth rate of existing methods when processing hydrological and meteorological data with complex nonlinear and multi-scale characteristics, this application proposes a hydrological time series prediction network and method based on feature extraction and guidance, adopts the encoder-decoder structure of Informer, and solves the above problems through multi-scale feature extraction in time domain, global feature extraction in frequency domain, dynamic feature fusion and guided sparse attention weight method.

[0007] The technical solution adopted in this application is: a hydrological time series prediction network based on feature extraction and guidance, characterized in that: an encoder-decoder model structure of Informer is adopted, and a feature extraction module is also included, and the output of the feature extraction module is used as the input of the encoder;

[0008] The feature extraction module includes a time domain feature extraction module, a frequency domain feature extraction module and a dynamic feature fusion module. The time domain feature extraction module is based on a U-shaped pyramid network structure and is used to extract multi-scale time domain features in hydrological time series data. The frequency domain feature extraction module extracts global frequency domain features in hydrological time series data through discrete cosine transform.

[0009] The dynamic feature fusion module is used to fuse time domain features and frequency domain features at the channel level and space level. Through the dynamic selection mechanism of global channel information and spatial information, the adaptive fusion of multi-scale features is realized.

[0010] The guided attention mechanism is used to replace the sparse attention mechanism in the original encoder. The guided attention mechanism optimizes the sparse probabilistic attention allocation by guiding the attention weights, thereby improving the ability to focus on key feature positions.

[0011] The time domain feature extraction module is based on a U-shaped network structure to extract and fuse the multi-scale features of the hydrological time series. The input is the preprocessed time series data. Through layer-by-layer average pooling operations, downsampling is performed to obtain features at different levels. The fusion and interaction of shallow and deep features are achieved through layer-by-layer feature splicing and linear changes. Shallow features are used to capture short-term local changes, while deep features focus more on long-term global trends.

[0012] The frequency domain feature extraction module uses discrete cosine transform to extract the frequency domain features of the input hydrological time series data, performs dynamic adaptation based on learnable frequency parameters, optimizes data allocation, and realizes sparse changes in frequency domain features through matrix changes, providing high-quality frequency domain feature representation for the subsequent dynamic feature fusion module.

[0013] The formula for discrete cosine transform is as follows:

[0014]

[0015] in is the frequency domain representation obtained through discrete cosine transform. z is a time series of length N, and the elements of the matrix are expressed as:

[0016]

[0017] where N is the length of the time series, k represents the frequency index, and n is the time index; ψ k is a learnable frequency parameter, whose initial value is set based on the standard DCT frequency and is then dynamically optimized during the training process; n f << N, n f is the effective frequency domain fraction obtained through discrete cosine transform. The retained frequency domain features are much fewer than the original length of the input sequence to achieve sparse representation of frequency domain features.

[0018] Channel-level fusion in the dynamic feature fusion module extracts importance weights through average pooling and 1×1 convolution to complete feature calibration in the channel dimension. Spatial-level fusion calculates global spatial weights through element-wise addition and the Sigmoid activation function, and outputs features with multiple scales and high and low dimensions.

[0019] The guided attention mechanism generates a global guidance vector by pooling or dimensionality reduction operations on the fused features output by the dynamic feature fusion module. The global guidance vector dynamically adjusts the weight distribution of the sparse self-attention mechanism through dot product weighting, which can effectively guide the attention to focus more on key feature positions.

[0020] A hydrological time series prediction method based on feature extraction and guidance, which adopts a hydrological time series prediction network based on feature extraction and guidance, includes the following steps:

[0021] S1: Data preprocessing: Preprocess the runoff data and meteorological data to ensure the symmetry and consistency of the data distribution; and use the preprocessed data as the dataset, and divide the dataset into a test set and a validation set;

[0022] S2: Feature extraction: Use the time domain feature extraction module and the frequency domain feature extraction module to extract the multi-scale features and periodic features of the time series data respectively;

[0023] S3: Feature fusion: Achieve the adaptive fusion of time domain features and frequency domain features through the dynamic feature fusion module to generate a unified feature representation;

[0024] S4: Model training: Train the model in a joint optimization manner, and introduce a hybrid loss function to minimize the error;

[0025] S5: Model prediction: Input the test set data and complete the prediction of short-term, medium-term and long-term time series through the improved Informer model.

[0026] In step S1, shift processing, distribution smoothing, and standardization operations are performed on the runoff data and meteorological data.

[0027] The parameters of the distribution smoothing are determined by the maximum likelihood estimation method to ensure that the smoothed data distribution is close to the normal distribution.

[0028] During the training process, the method of verifying round by round is adopted, and the optimal model parameters are selected through the validation set; the mean square error, root mean square error, mean absolute error, mean absolute percentage error, and mean square percentage error indicators are used to evaluate the model.

[0029] The beneficial effects of this application compared with the prior art are as follows:

[0030] (1) The fusion of time-domain features and frequency-domain features can improve the prediction accuracy of the model;

[0031] (2) The dynamic feature fusion can enhance the robustness of the model;

[0032] (3) The guided attention mechanism can optimize the capture of key features. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The following further describes this application with reference to the drawings:

[0034] Figure 1 It is a schematic structural diagram of the improved network provided by the embodiment of this application;

[0035] Figure 2 It is a schematic structural diagram of TimeFlow-Net provided by the embodiment of this application;

[0036] Figure 3 It is a schematic structural diagram of FreqScope provided by the embodiment of this application;

[0037] Figure 4 It is a schematic structural diagram of FusionBridge provided by the embodiment of this application;

[0038] Figure 5 It is a schematic structural diagram of GuideFocus Attention provided by the embodiment of this application. DETAILED DESCRIPTION OF THE INVENTION

[0039] As Figures 1 to 5As shown, the present application provides a hydrological time series prediction network based on feature extraction and guidance, which adopts the encoder-decoder structure of Informer, specifically including a time domain feature extraction module (Time Domain FlowNetwork, TimeFlow-Net), a frequency domain feature extraction module (Frequency Spectrum Scope, FreqScope), a dynamic feature fusion module (FusionBridge) and a guided attention mechanism (GuideFocus Attention), wherein the time domain feature extraction module extracts multi-scale time domain features in hydrological time series data based on a U-shaped pyramid network structure to enhance the time domain modeling capability of the model; the frequency domain feature extraction module extracts global frequency domain features in hydrological time series data through a custom discrete cosine transform (CDCT) to optimize the model's non-stationary sequence modeling capability; the dynamic feature fusion module realizes adaptive fusion of multi-scale features through a dynamic selection mechanism of global channel information and spatial information, thereby improving the model's ability to express hydrological time series features; the guided attention mechanism module optimizes sparse probabilistic attention allocation by guiding attention weights to enhance the ability to focus attention on key feature positions.

[0040] In order to effectively extract multi-scale time-domain features from time series data, this application introduces the TimeFlow-Net module and integrates it into the overall model architecture as a time-domain feature extractor. The time-domain feature extraction module is based on the U-shaped network structure. Combined with the characteristics of time series data, the original U-Net is adapted and improved. It mainly includes two core parts: the time feature pyramid network (Time series FPN) and the multi-scale feature fusion module. The TimeFlow-Net module is used to extract and fuse the multi-scale features of hydrological time series. The input is pre-processed time series data. Through layer-by-layer average pooling operations, downsampling is performed to obtain features at different levels, and the fusion and interaction of shallow features and deep features are realized through layer-by-layer feature splicing and linear changes. Shallow features are used to capture short-term local changes, and deep features pay more attention to long-term global trends.

[0041] like Figure 2 As shown, the input of the time feature pyramid network is the original time series data x∈R B×T×C , where B is the batch size, T is the time step, C is the number of channels, and R is a real number. Through the layer-by-layer average pooling operation, downsampling obtains multiple features of different scales X = {x1, x2, x3, ..., x n}, where x1 is the original input, x i is the downsampling result of the i-th layer, and the calculation formula is as follows:

[0042] x i= AvgPool(x i-1 ), for i = 2, 3, ..., n;

[0043] The parameters of average pooling are set as: kernel size kernel_size = 3, stride stride = 2, padding padding = 0.

[0044] The formula for calculating the feature length is:

[0045]

[0046] The multi-scale time-domain features X = {x1, x2, x3, …, x n} extracted in the time feature pyramid network need to be further fused to make full use of the feature information at different scales. Feature fusion and reconstruction are achieved through layer-by-layer feature concatenation and linear transformation.

[0047] Feature fusion: Let y i represent the feature of the i-th layer. The fusion process is as follows:

[0048] y’ i-1 = Linear(cat(y’ i , y i-1 ), for i = 2, 3, …, n;

[0049] where cat represents the feature concatenation operation, Linear is the fully connected layer, y’ i-1 is the prediction result of the previous stage i - 1, y’ i is the prediction result of the current stage i, and y i-1 is the original output of the previous stage.

[0050] For each layer of feature y i , the length calculation is the same as that of the input feature x, and the dimension of the output feature is calculated based on the pooling operation.

[0051]

[0052] Through this layer-by-layer fusion, the shallow and deep features interact fully, thus enhancing the robustness of feature expression.

[0053] The multi-scale fused feature y’ i is used as the output of the TimeFlow-Net module, providing a high-quality time-domain feature representation and serving as the input for the subsequent dynamic feature fusion module FusionBridg and prediction tasks.

[0054] The FreqScope module uses a custom discrete cosine transform (CDCT) to extract the frequency-domain features of the input hydrological time series data. It can be dynamically adapted based on the learned frequency parameters to optimize data allocation, and realizes the sparse transformation of frequency-domain features through matrix transformation, providing a high-quality frequency-domain feature representation for the subsequent dynamic feature fusion module.

[0055] As Figure 3 shown, the FreqScope module first normalizes the input data:

[0056]

[0057] where μ and σ are the mean and standard deviation respectively. Subsequently, the time series is divided into continuous small blocks through the block technology to form the basic input units.

[0058] The core idea of CDCT is to introduce learnable frequency parameters so that the transformation basis can better adapt to the actual distribution of the data, thereby improving the sparsity and expressive power of the frequency-domain representation.

[0059] CDCT is defined as follows:

[0060]

[0061] where is the frequency-domain representation obtained through the discrete cosine transform, z is the time series of length N, and the elements of the matrix are expressed as:

[0062]

[0063] where N is the time series length, k represents the frequency index, n is the time index; ψ k is the learnable frequency parameter, whose initial value is set based on the standard DCT frequency and is then dynamically optimized during the training process; n f << N, n f is the effective frequency-domain fraction obtained through the discrete cosine transform. The retained frequency-domain features are much fewer than the original length of the input sequence to achieve the sparse representation of frequency-domain features.

[0064] The dynamic feature fusion module performs channel-level and spatial-level fusion on the time-domain features and frequency-domain features. Channel-level fusion extracts the importance weights through average pooling 1×1 convolution to complete the feature calibration in the channel dimension. Spatial-level fusion calculates the global spatial weights through element-wise addition and the Sigmoid activation function, thereby outputting features with multi-scale and high-low dimensions, so as to improve the model's representation ability for hydrological time series.

[0065] As Figure 4As shown, the FusionBridge module realizes the adaptive fusion of multi-scale features through the dynamic selection mechanism of global channel information and spatial information. First, the time-domain and frequency-domain feature maps are concatenated. and The channel-level importance weight w is extracted through average pooling and 1×1 convolutional layers. ch Furthermore, the feature calibration in the channel dimension is completed.

[0066]

[0067] Next, the channel information w ch is used to perform weighted sum and selection on the feature map to obtain the feature map F l .

[0068] F l = Conv1(w ch ⊙ [F l 1 ; F l 2 ).

[0069] In addition, the FusionBridge module also extracts the global spatial weight w sp through element-wise addition and the Sigmoid activation function for further calibrating the spatial dimension of the feature map;

[0070]

[0071] Finally, the global spatial weight w sp is used to adjust the feature map to obtain the spatially adjusted feature map

[0072] Output feature The fused local and global information in the frequency domain improves the model's representation ability and prediction accuracy for time series data.

[0073] The guided attention mechanism generates a global guidance vector through pooling or dimensionality reduction operations on the fused features output by the dynamic feature fusion module. This vector dynamically adjusts the weight distribution of the sparse self-attention mechanism through dot product weighting, effectively guiding the attention to focus more on key feature positions.

[0074] Such as Figure 5As shown, in order to further improve the performance of the ProbSparse Attention mechanism, this application proposes a sparse attention mechanism combined with fusion feature guidance. By guiding the attention weights (the time-domain and frequency-domain features fused through FusionBridge) to provide a more intelligent guidance signal for attention calculation, not only can the computational efficiency be improved by optimizing the attention distribution, but also the attention focusing ability of the model can be improved, thereby enhancing the accuracy when processing complex input data.

[0075] Suppose there is a query matrix key matrix and value matrix At the same time, this embodiment also defines a fusion feature matrix obtained through a dynamic feature fusion module (FusionBridge) where is the real number field, L Q is the length of the query matrix Q, L K is the length of the key matrix K, L V is the length of the value matrix V, and d is the dimension of the input.

[0076] First, perform pooling or dimensionality reduction operations on the fusion feature to obtain a guidance vector G guide , which has the same dimension as the query vector Q:

[0077]

[0078] Among them, the Poolin operation can be mean pooling, max pooling or random pooling. The pooled guidance vector Gguide represents the importance of global and local information in the fusion feature.

[0079] In the ProbAttentio mechanism, the original attention score matrix M is obtained by calculating the dot product between the query Q and the key K, that is:

[0080] M = Q · K T ;

[0081] Through the guidance vector G guide , the original attention scores are weighted and adjusted to dynamically control the distribution of attention weights. The adjusted attention score matrix is:

[0082]

[0083] Among them, λ is a learnable hyperparameter used to control the contribution of the guidance vector to the attention scores. The introduction of

[0084] Next, by applying the Softmax function to the adjusted attention score matrix M adjust to calculate the new attention weight distribution and applying it to the value matrix V:

[0085] A adjust = Softmax(M adjust )V;

[0086] where A adjust is the new attention weight distribution.

[0087] This process completes the update of the sparse attention, and the finally obtained attention weight distribution A adjust reflects the query-key relationship under the guidance of the fused features.

[0088] This application also proposes a hydrological time series prediction method based on feature extraction and guidance, which mainly includes the following steps:

[0089] S1: Data preprocessing: Perform shifting, distribution smoothing (Box-Cox transformation), and standardization (Z-Score) operations on the runoff data and meteorological data to ensure the symmetry and consistency of the data distribution; the preprocessed data is used as the dataset, and the dataset is divided into a test set and a validation set.

[0090] The Box-Cox transformation formula is as follows:

[0091]

[0092] In the formula: y(λ) is the transformed variable value; y is the original variable value, taken from the processed dataset (all values have been shifted to be positive); λ is the optimal transformation parameter determined by the maximum likelihood estimation method (MLE) to make the distribution of the transformed data as close to the normal distribution as possible.

[0093] The Z-Score formula is as follows:

[0094]

[0095] In the formula: z is the standardized variable value; x is the original variable value, taken from the processed dataset (which has been smoothed by the Box-Cox transformation); μ is the mean of each variable, calculated based on the training set; σ is the standard deviation of each variable, calculated based on the training set.

[0096] S2: Feature extraction: Use the time-domain feature extraction module and the frequency-domain feature extraction module to extract the multi-scale features and periodic features of the time series data respectively.

[0097] S3: Feature Fusion: Achieve the adaptive fusion of time-domain features and frequency-domain features through a dynamic feature fusion module to generate a unified feature representation.

[0098] S4: Model Training: Train the model in a joint optimization manner, introducing a hybrid loss function (MSE and MAE) to minimize the error;

[0099] The formula for the hybrid loss function L is as follows:

[0100] L = α·MSE + β·MAE;

[0101] where α and β represent weights respectively.

[0102] S5: Model Prediction: Input the test set data and complete the prediction of short-term, medium-term, and long-term time series through the improved Informer model.

[0103] During the training process, the method of verification round by round is adopted, and the model parameters with the optimal performance are selected through the validation set. The model evaluation uses multiple metrics, including mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and mean squared percentage error (MSPE). The definitions of each metric are as follows:

[0104]

[0105]

[0106] where n is the total number of samples, Y i is the observed value, and Y i is the predicted value.

[0107] The prediction task covers short-term, medium-term, and long-term time series; the model optimizes the modeling ability of long-term dependencies through dynamic feature fusion and a guided attention mechanism, and shows significant advantages in the growth rate of prediction errors.

[0108] To verify the advantages of the model proposed in this application in hydrological time series prediction, this application selects six networks for comparative analysis. Table 1 is the comparison table of evaluation metrics for each model.

[0109] Table 1 Comparison Table of Model Evaluation Metrics

[0110]

[0111] From the experimental results, it can be seen that the model (Ours) shows significant advantages at all prediction lengths. Especially in key indicators such as MSE, RMSE, and MAE, it outperforms other models. Its error control ability and stability are more prominent as the prediction step increases, reflecting the efficient modeling ability of the model in complex hydrological time series prediction tasks.

[0112] Judging from the experimental data, this model is lower than other comparison models in all evaluation indicators. For example, the MSE is 0.537, significantly lower than that of Transformer (0.617) and Reformer (0.612), showing high adaptability to hydrological time series prediction tasks. Especially in terms of the MAE and RMSE indicators, this model is 0.544 and 0.732 respectively, which is more accurate than other models, indicating that it has significant advantages in error control.

[0113] Compared with models such as Transformer and Reformer, this model shows significant advantages in capturing non-linear time series features. Taking MAPE as an example, this model is only 4.11%, while both Transformer and Reformer are 4.73%, indicating that this model can capture key changes in the sequence more accurately. This performance is mainly attributed to the synergistic effect of the time-domain feature extraction module (TimeFlow-Net) and the frequency-domain feature extraction module (FreqScope):

[0114] The TimeFlow-Net module extracts multi-scale time-domain features through a pyramid structure, effectively enhancing the modeling ability of local fluctuations in time series.

[0115] The FreqScope module optimizes the extraction ability of global frequency-domain features through custom discrete cosine transform (CDCT), thereby improving the processing effect of non-stationary data.

[0116] From the experimental results, it can be seen that models such as Autoformer and FEDformer improve their adaptability to complex time series through decomposition mechanisms, but the error level is still slightly higher than that of this model. For example, the MSE of Autoformer and FEDformer are 0.601 and 0.596 respectively, while this model is 0.537, with obvious performance improvement. This indicates that this model performs more stably in complex feature interaction and long-time series modeling.

[0117] This stability benefits from the dynamic feature fusion module (FusionBridge), which achieves efficient fusion of time-domain and frequency-domain features through channel- and space-level dynamic feature selection, significantly enhancing the model's ability to represent multivariate data. Additionally, the introduction of the guided attention mechanism (GuideFocus Attention) enables the model to focus more on key features in the time series by optimizing the sparse attention weight distribution, further improving the modeling ability for complex patterns.

[0118] In the comparison with lightweight models such as PatchTST, it can be found that although the short-term prediction performance of PatchTST is close to that of this model, for example, the MSE is 0.575, the overall error level is still higher than that of this model. This indicates that PatchTST has certain ability in capturing short-term features, but its modeling ability and generalization performance for complex time series features are inferior to this model. In contrast, the innovative module design of this model (such as time-domain and frequency-domain feature extraction modules and dynamic fusion mechanisms) endows the model with stronger stability and adaptability.

[0119] In summary, the model proposed in this application can better predict multivariate time series in meteorology, extract time features at different scales, improve the expression ability of frequency-domain features, and guide the optimization of attention weights. The synergistic effect of these modules has obvious advantages compared with other time series models.

[0120] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of this application, rather than to limit it; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A hydrological time series prediction network based on feature extraction and guidance, characterized in that: The encoder-decoder model structure of Informer is adopted, and a feature extraction module is also included. The output of the feature extraction module serves as the input of the encoder. The feature extraction module includes a time-domain feature extraction module, a frequency-domain feature extraction module, and a dynamic feature fusion module. The time-domain feature extraction module is based on the U-shaped pyramid network structure and is used to extract multi-scale time-domain features in hydrological time series data. The frequency-domain feature extraction module extracts global frequency-domain features in hydrological time series data through discrete cosine transform. The dynamic feature fusion module is used to perform channel-level and spatial-level fusion on time-domain features and frequency-domain features. Through the dynamic selection mechanism of global channel information and spatial information, the adaptive fusion of multi-scale features is achieved. The guided attention mechanism is adopted to replace the sparse attention mechanism in the original encoder. The guided attention mechanism optimizes the sparse probability attention distribution by guiding attention weights, enhancing the focusing ability of attention on key feature positions.

2. The hydrological time series prediction network based on feature extraction and guidance according to claim 1, characterized in that: Based on the U-shaped network structure, the time-domain feature extraction module extracts and fuses multi-scale features of hydrological time series. The input is preprocessed time series data. Through layer-by-layer average pooling operations, different levels of features are downsampled, and the fusion and interaction of shallow features and deep features are achieved through layer-by-layer feature concatenation and linear transformation. Shallow features are used to capture short-term local changes, while deep features pay more attention to long-term global trends.

3. The hydrological time series prediction network based on feature extraction and guidance according to claim 1, wherein: The frequency-domain feature extraction module uses discrete cosine transform to extract the frequency-domain features of the input hydrological time series data, performs dynamic adaptation based on learnable frequency parameters, optimizes data allocation, and realizes the sparse change of frequency-domain features through matrix transformation, providing high-quality frequency-domain feature representations for the subsequent dynamic feature fusion module.

4. The hydrological time series prediction network based on feature extraction and guidance according to claim 3, characterized in that: The formula for discrete cosine transform is as follows: wherein is the frequency domain representation obtained by discrete cosine transform, z is a time series of length N, and the elements of the matrix are expressed as: where N is the length of the time series, k represents the frequency index, and n is the time index; ψ k is a learnable frequency parameter, whose initial value is set based on the standard DCT frequencies and is then dynamically optimized during training; n f << N, n f is the effective frequency-domain fraction obtained by discrete cosine transform.

5. A hydrological time series prediction network based on feature extraction and guidance according to claim 1, characterized in that: In the dynamic feature fusion module, channel-level fusion extracts importance weights through average pooling 1×1 convolution to complete feature calibration in the channel dimension. Spatial-level fusion calculates global spatial weights through element-wise addition and the Sigmoid activation function, and outputs features with both multi-scale and high-low dimension.

6. The hydrological time series prediction network based on feature extraction and guidance according to claim 1, characterized in that: The guided attention mechanism generates a global guiding vector through pooling or dimensionality reduction operations on the fused features output by the dynamic feature fusion module. The global guiding vector dynamically adjusts the weight distribution of the sparse self-attention mechanism through dot product weighting, effectively guiding attention to focus more on key feature positions.

7. A hydrological time series prediction method based on feature extraction and guidance, characterized in that: The hydrological time series prediction network based on feature extraction and guidance as described in any one of claims 1-6 is adopted, including the following steps: S1: Data preprocessing: Preprocess the runoff data and meteorological data to ensure the symmetry and consistency of data distribution; and The preprocessed data serves as the data set, and the data set is divided into a test set and a validation set. S2: Feature extraction: Use the time-domain feature extraction module and the frequency-domain feature extraction module to extract multi-scale features and periodic features of time series data respectively. S3: Feature fusion: Achieve the adaptive fusion of time-domain features and frequency-domain features through the dynamic feature fusion module to generate a unified feature representation. S4: Model training: Train the model in a joint optimization manner, introducing a hybrid loss function to minimize the error. S5: Model Prediction: Input the test set data, and complete the prediction of short-term, medium-term, and long-term time series through the improved Informer model.

8. A method for predicting hydrological time series based on feature extraction and guidance according to claim 7, characterized in that: In step S1, shift processing, distribution smoothing, and standardization operations are performed on the runoff data and meteorological data.

9. A method for predicting hydrological time series based on feature extraction and guidance according to claim 8, characterized in that: The parameters of distribution smoothing are determined by the maximum likelihood estimation method to ensure that the distribution of the smoothed data is close to the normal distribution.

10. A hydrological time series prediction method based on feature extraction and guidance according to claim 8, characterized in that: During the training process, the method of verification round by round is adopted, and the model parameters with the optimal performance are selected through the validation set; the mean square error, root mean square error, mean absolute error, mean absolute percentage error, and mean square percentage error indicators are used to evaluate the model.

Images