Multi-scale memory and deep learning-based flood forecasting method and system for areas lacking data

Abstract

The invention belongs to the related technical field of hydrological forecasting, and discloses a flood forecasting method and system for an area lacking data based on multi-scale memory and deep learning, and the method comprises the steps: collecting observation value data of flood runoff of a plurality of hydrological stations; for the missing value, acquiring a time correlation quantity according to the observation value information with similar time, and acquiring a space correlation quantity according to the observation value information with similar space, so as to acquire a single-point spatial-temporal characteristic of the missing value; acquiring local spatio-temporal features including all hydrometric site features; inputting the local spatial-temporal characteristics into a multi-scale memory network storing multiple scale observation value data, and outputting generated value data to realize runoff embedding; and flood forecasting is carried out based on the runoff data after runoff embedding. According to the method, missing value feature extraction is carried out by comprehensively considering the hydrological site time-space relationship, and the multi-scale historical features and the local time-space features are fused through the multi-scale memory network structure, so that the current feature embedding can be enhanced, and the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field related to hydrological forecasting, and more specifically, relates to a flood forecasting method and system for data-scarce areas based on multi-scale memory and deep learning. Background Art

[0002] Missing time series data is a common problem in flood forecasting modeling. Missing data makes it difficult to capture the evolution of the entire sequence and prevents the prediction model from fully extracting valuable features from the sequence, significantly reducing forecast accuracy. Hydrological data used in flood forecasting is a time series data that meets these characteristics. There are two methods for handling missing data, depending on how missing values ​​are handled.

[0003] Deletion: This method directly deletes all feature variables with missing values. While this method solves the problem of missing values, it also reduces the amount of data, resulting in a significant loss of information. This is detrimental to predictive models that exploit the characteristics and evolutionary patterns of time series, often failing to achieve good predictive performance. Interpolation: This method does not delete any sample data, but instead uses various methods to fill in missing data, ensuring data integrity. The development of missing value imputation technology for hydrological data can be divided into three stages.

[0004] In the early days, traditional statistical methods dominated, with classic techniques such as linear interpolation, Lagrange interpolation, and polynomial interpolation long dominating. These methods estimate missing values ​​by establishing linear or polynomial relationships between adjacent data points, and are well-suited for scenarios with low data volatility and limited missing values. With technological advances, methods such as spline interpolation and kriging interpolation have gradually gained application, achieving more accurate interpolation by constructing continuous functions around data points. However, these methods often struggle to effectively capture the complex distributional characteristics of highly heterogeneous data. The introduction of time series models has opened up new avenues for missing value imputation. Time series analysis methods, such as ARIMA models and seasonal models, predict missing values ​​by exploiting historical patterns in data. While these methods offer the advantage of ease of implementation, their accuracy can be limited when dealing with complex time series with multi-scale features.

[0005] In recent years, breakthroughs in artificial intelligence (AI) technology have driven innovation in the field of hydrological data interpolation. Machine learning algorithms have demonstrated powerful nonlinear modeling capabilities. For example, researchers have constructed an interpolation model for annual precipitation data in the Three Gorges region based on support vector machines (SVMs). Others have used random forest methods to achieve superior interpolation results compared to kriging on a synthetic dataset from Spain. The introduction of deep neural network models has also brought about a paradigm shift. Researchers have proposed an improved recurrent neural network (RNN) architecture to successfully address noise interference in automatic calibration signal (ACS) data. Others have combined long short-term memory (LSTM) networks with Hermite spline interpolation (CHS) to effectively recover high-resolution data from vessel monitoring systems (VMS). These methods significantly improve the accuracy of missing value interpolation by capturing long-range dependencies in the data.

[0006] However, flood runoff data not only has temporal correlation, but also spatial correlation. Spatially connected river basins will also have an impact on flood runoff. Existing flood prediction models mainly deal with missing values ​​from the time dimension and do not fully consider the complex temporal and spatial characteristics of flood runoff data. The prediction accuracy needs to be further improved. Summary of the Invention

[0007] In response to the above-mentioned defects or improvement needs of the existing technology, the present invention provides a flood forecasting method and system for data-scarce areas based on multi-scale memory and deep learning, which is used to solve the problem that the existing flood prediction model mainly processes missing values ​​from the time dimension, does not fully consider the complex spatiotemporal characteristics of flood runoff data, and the prediction accuracy needs to be further improved.

[0008] To achieve the above objectives, according to one aspect of the present invention, a flood forecasting method for data-scarce areas based on multi-scale memory and deep learning is provided, comprising: Runoff collection: Collect observation data of flood runoff at multiple hydrological stations in the target hydrological station; Missing value feature extraction: For missing values ​​in the observation data, the time correlation of the missing value is obtained based on the observation information close to the time of the missing value, and the spatial correlation of the missing value is obtained based on the observation information close to the hydrological station where the missing value is located; the single-point spatiotemporal characteristics of the missing value are obtained based on the time correlation and spatial correlation; and then the local spatiotemporal characteristics corresponding to the moment of the missing value, including the single-point spatiotemporal characteristics corresponding to all hydrological stations, are obtained; Runoff generated value embedding: The local spatiotemporal features are input into a multi-scale memory network to generate a query vector, and based on the query vector and the memory data in the multi-scale memory network, generated value data including flood runoff at all hydrological stations corresponding to the time of the missing value is output to achieve runoff embedding; wherein the memory data is obtained by inputting observation data at multiple scales into the multi-scale memory network; Flood forecasting: Based on the runoff data after the runoff generation value is embedded, flood forecasting is performed using a prediction model.

[0009] According to the flood forecasting method for data-deficient areas based on multi-scale memory and deep learning provided by the present invention, obtaining the time-related quantity of the missing value in the missing value feature extraction specifically includes: The empirical time mean is obtained by using multiple observations of the hydrological station where the missing value is located before the time corresponding to the missing value. ; Get the last time observation value based on the closest and complete observation value before the corresponding time of the missing value at the hydrological station where the missing value is located and the time interval between the two ; Accordingly, obtaining the spatial correlation of the missing value specifically includes: The empirical spatial mean is obtained by using the observation values ​​corresponding to multiple hydrological stations close to the hydrological station where the missing value is located. ; Obtain the nearest spatial observation value based on the observation value of the hydrological station that is closest to and complete to the hydrological station where the missing value is located and the spatial distance between the two .

[0010] According to the flood forecasting method for data-deficient areas based on multi-scale memory and deep learning provided by the present invention, the empirical time mean The formula is as follows: ; Empirical space mean The formula is as follows: ; The single-point spatiotemporal characteristics of the missing value As shown in the following formula: ; Where, For the i The first hydrological station t The observed value at the time; is a mask, 0 indicates that the corresponding observation is missing, and 1 indicates that the corresponding observation is complete; L Indicates taking t Before the moment LObservations are used to obtain the empirical time mean; S Indicates taking the hydrological station close to the missing value S The empirical spatial mean of the observation values ​​corresponding to the hydrological stations is obtained; is the time decay rate; is the spatial attenuation rate.

[0011] According to the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning provided by the present invention, the spatiotemporal attenuation rate is modeled as: ; in, , For hydrological stations i The corresponding time decay parameter controls the influence of time distance; ; in, , For hydrological stations i The corresponding spatial attenuation parameter controls the influence of spatial distance.

[0012] According to the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning provided by the present invention, the observation data of multiple scales embedded in the runoff generation value include multiple scales of daily scale periodic data, weekly scale periodic data, monthly scale periodic data, ten-day scale periodic data and annual scale periodic data.

[0013] According to the flood forecasting method for data-deficient areas based on multi-scale memory and deep learning provided by the present invention, the observation data of multiple scales embedded in the runoff generation value include daily scale periodic data, weekly scale periodic data, and monthly scale periodic data. The observation data of multiple scales are specifically as follows: The data at each moment in the daily periodic data include the observation data of multiple days before that moment; The data at each moment in the weekly periodic data include the observation data of multiple days before and after the moment, centered on the day corresponding to the moment. The data at each moment in the monthly periodic data include the observation data for multiple days before and after the moment, centered on the day corresponding to the moment, in the multiple months before the moment.

[0014] According to the flood forecasting method for data-deficient areas based on multi-scale memory and deep learning provided by the present invention, the generated value data of flood runoff at all hydrological stations corresponding to the time of the missing value is output in the runoff generated value embedding, as follows: The multi-scale memory network calculates the query vector through inner product and Softmax function With each input memory vector Attention score : ; The attention score reflects the similarity between historical observations and queries. A higher score indicates that the pattern is more relevant to the current missing value context. The memory response vector By outputting the memory vector The weighted sum of is generated: ; Finally, the output enhanced runoff embedding is as follows: ; in, for t Generated value data of runoff at each moment; 、 As parameters, Indicates splicing, represents the number of input memory vectors and the number of output memory vectors, wherein the input memory vectors and the output memory vectors form the memory data.

[0015] According to the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning provided by the present invention, the prediction model in flood forecasting adopts a hybrid model combining a graph convolutional network and a long short-term memory network, wherein the graph convolutional network is used to extract the spatial features of the runoff data after the generated value is embedded, and the long short-term memory network is used to extract the temporal features based on the spatial features and output the forecast results.

[0016] According to the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning provided by the present invention, the input of the graph convolutional network is as follows: Establishing a network graph of the target hydrological station including a plurality of hydrological sites, wherein the network graph includes nodes consisting of the plurality of hydrological sites and edges connecting the nodes with flow relationships, wherein each node has location information and runoff data information; Based on the network graph, an adjacency matrix and a node feature matrix are established as inputs of the graph convolutional network.

[0017] According to another aspect of the present invention, a flood forecasting system for data-scarce areas based on multi-scale memory and deep learning is provided. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it executes any one of the above-mentioned flood forecasting methods for data-scarce areas based on multi-scale memory and deep learning.

[0018] In general, compared with the prior art, the above technical solutions conceived by the present invention provide a flood forecasting method and system for data-scarce areas based on multi-scale memory and deep learning: 1. Considering the strong spatial correlation between hydrological stations can provide more information for missing value processing. It is proposed to comprehensively consider the spatiotemporal relationship between hydrological stations to extract missing value features. A multi-scale memory network structure is used to store multi-scale global historical patterns. By embedding multi-scale historical observation data into memory, it can provide supplementary information for local spatiotemporal features. By fusing multi-scale historical features with local spatiotemporal features, it can not only eliminate the negative impact of missing values ​​but also enhance the embedding of current features, thereby providing more useful information for data-scarce areas and improving prediction accuracy. 2. For missing values, the empirical temporal mean, the last observed value, the time interval between the last observed value, the empirical spatial mean, the nearest observed value, and the spatial distance between the last observed value are extracted as relevant quantities to construct a single-point spatiotemporal feature of the missing value. This, in turn, forms a local spatiotemporal feature. This integrates temporal and spatial information, improving the performance and reliability of the generated value-enhanced embedding. The constructed exponential spatiotemporal decay rate can better and more reasonably balance the effects of various relevant quantities, enhancing intelligence and flexibility. 3. A hybrid model combining a graph convolutional network (GCN) and a long short-term memory (LSTM) network was constructed to perform flood forecasting based on the enhanced embedded runoff data. The enriched missing data values ​​and the graph adjacency matrix were input into the model. The GCN-LSTM framework was used as a feature extractor, and the final prediction results were generated through fully connected layers. This model can simultaneously capture the spatial dependencies and temporal dynamics of the data, which helps improve prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning provided by the present invention.

[0020] Figure 2 This is a logical diagram of a multi-scale memory network in an embodiment of the present invention that uses local spatiotemporal features and global historical patterns to enrich runoff embedding.

[0021] Figure 3 Schematic diagram of the GCN-LSTM model structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0023] See also Figure 1This embodiment provides a flood forecasting method for data-scarce areas based on multi-scale memory and deep learning, the method comprising: Runoff collection: Collect observation data of flood runoff from multiple hydrological stations in the target hydrological station; Missing value feature extraction: For missing values ​​in the observation data, the time correlation of the missing value is obtained based on the observation information close to the time of the missing value, and the spatial correlation of the missing value is obtained based on the observation information close to the hydrological station where the missing value is located; the single-point spatiotemporal characteristics of the missing value are obtained based on the time correlation and spatial correlation; and then the local spatiotemporal characteristics corresponding to the moment of the missing value, including the single-point spatiotemporal characteristics corresponding to all hydrological stations, are obtained; Runoff generated value embedding: The local spatiotemporal features are input into a multi-scale memory network to generate a query vector, and based on the query vector and the memory data in the multi-scale memory network, generated value data including flood runoff at all hydrological stations corresponding to the time of the missing value is output to achieve runoff embedding; wherein the memory data is obtained by inputting observation data at multiple scales into the multi-scale memory network; Flood forecasting: Based on the runoff data after the runoff generation value is embedded, flood forecasting is performed using a prediction model.

[0024] The technical problem to be solved by the embodiments of the present invention is to overcome the defect that the existing data-deficient basin flood prediction technology does not take into account the strong spatial correlation between hydrological stations, and to provide a data-deficient basin flood prediction technology based on the spatiotemporal relationship of stations, so as to provide more useful information for data-deficient areas and thus improve the prediction accuracy.

[0025] Define the target hydrological station network as ,in express N A collection of hydrological stations, Indicates the connection to these sites E A collection of edges, where a connecting edge is set between two sites with a circulation relationship. The observation data of each site at each moment contains F characteristics such as runoff, rainfall, evapotranspiration, etc. Use Indicates that all sites are All eigenvalues ​​within a time slice. Indicates time t All characteristic values ​​of all sites, and yes Middle i The eigenvalue vector of each site. At the same time, the mask sequence is defined ,in ,here Indicates the i Sites at timet The eigenvalue of is missing. If missing, otherwise , the goal of this embodiment is to build a model , which can be used with incomplete runoff sequences and hydrological station network As input, predict the future Runoff data for time steps: .

[0026] In some specific embodiments, considering that small watersheds have complex terrain and fast confluence speed, traditional flood forecasting methods are difficult to apply effectively in data-scarce areas due to the lack of data. This embodiment aims to use the spatial data processing capabilities of graph neural networks and the temporal processing capabilities of recurrent neural networks combined with memory networks that enrich missing data to achieve accurate flood forecasting in such areas. This specific embodiment is divided into the following steps: first, missing data is enriched by using a multi-scale memory network to consider the spatiotemporal correlation between hydrological stations. Secondly, a graph convolutional network (GCN) is used to capture the spatial correlation between node attributes, and a long short-term memory network (LSTM) is used to simulate the temporal correlation based on historical data. This combined approach can effectively simulate the spatiotemporal correlation observed by monitoring stations. The GCN-LSTM framework is used as a feature extractor and finally a fully connected layer is used to generate the final prediction result.

[0027] First, a multi-scale memory network is used to enrich the runoff data: To enrich the input data with the complex spatiotemporal relationships between hydrological stations, a memory network is employed. The key idea behind the memory network is to learn from historical memory components that preserve long-range, multi-scale patterns, namely, diurnal, weekly, and monthly dependencies. The scale range depends on the characteristics of the data. Specifically, local spatiotemporal features are extracted as the key to querying the memory component. These local spatiotemporal features are derived from the station with missing runoff values, and spatial features are derived from the stations adjacent to the missing values. Weighted historical long-term patterns, i.e., memory data, are combined with local statistical features to eliminate the side effects of missing values. The local-global features are then output as a rich flow embedding.

[0028] refer to Figure 2 First, a feature extraction module can be constructed to extract missing value features based on the original observation data, and the contextual information of the time series observations can be used to extract spatiotemporal features. Unlike previous studies, this embodiment considers both temporal and spatial aspects when generating the following statistical features for each timestamp: First, obtain the empirical time mean through multiple observations of the hydrological station where the missing value is located before the corresponding time of the missing value ; Get the last time observation value based on the closest and complete observation value before the corresponding time of the missing value at the hydrological station where the missing value is located and the time interval between the two ; to obtain the time-dependent amount of the missing value.

[0029] Experience time mean : missing values The empirical time mean through time t Previous L Historical observations are calculated to reflect recent status and provide The specific formula is as follows: ; Represents the time step Time Station Is the observation data missing (1 is complete, 0 is missing) (using as a mask to create missing values ​​for the complete dataset).

[0030] Last time observation: Based on the assumption that any missing value will more or less inherit the information of the last non-missing observation. That is, the time adjacent values ​​are kept close to the current missing value. Assuming missing values The information of the last valid observation value can be obtained from Inheritance, the time interval between the two is (For example, taking the number of time steps between them).

[0031] Second, the empirical spatial mean is obtained by using the observation values ​​corresponding to multiple hydrological stations close to the hydrological station where the missing value is located. ; Obtain the nearest spatial observation value based on the observation value of the hydrological station that is closest to and complete to the hydrological station where the missing value is located and the spatial distance between the two ; to obtain the spatial correlation of the missing values.

[0032] Empirical space mean : For each missing value , using hydrological stations i Nearby S Sample distance stations i Recent S The observation values ​​of each station are used to form a sample to construct its empirical spatial mean to reflect the local runoff state. The specific formula is: ; Indicates the neighboring site s at time step t The mask value of .

[0033] Nearest spatial observation: Usually, the state of a graph node is similar to that of its neighboring nodes, and neighboring sites share similar runoff conditions. Definition For nodes i At time step t The closest and complete spatial observation value (such as the node with the closest geographical location) has a spatial distance of (e.g. Euclidean distance).

[0034] Time-space decay rate: Based on time-space distance, design a trainable time decay rate and spatial attenuation rate , used to dynamically weigh the contribution of the last observation and the mean. Generally speaking, when and When they are younger, they are more likely to believe and When the space / time distance becomes larger, the space / time average will be more representative. According to this assumption, the space / time decay rate Modeled as: ; in, , For hydrological stations i The corresponding time decay parameter controls the influence of time distance; ; in, , For hydrological stations i The corresponding spatial attenuation parameter controls the influence of spatial distance. The time attenuation parameter and the spatial attenuation parameter can be determined through training. The exponential negative rectifier is selected to make the attenuation rate and It decreases monotonically in the range of 0 to 1.

[0035] Taking into account trainable decay, the proposed model combines spatial / temporal estimation to define missing values Single-point spatiotemporal characteristics As shown in the following formula: ; Where, For the i The first hydrological station t The observed value at the time; is a mask, 0 indicates that the corresponding observation is missing, and 1 indicates that the corresponding observation is complete; LIndicates taking t Before the moment L Observations are used to obtain the empirical time mean; S Indicates taking the hydrological station close to the missing value S The empirical spatial mean of the observation values ​​corresponding to the hydrological stations is obtained; is the time decay rate; is the spatial decay rate. Therefore, for missing values , we can get the local spatiotemporal characteristics of the corresponding moment .

[0036] Then, a multi-scale memory network structure is constructed, and the global historical pattern plays a key role in building rich runoff embedding data. By embedding multi-scale (daily, weekly, monthly) historical observation data into memory, local spatiotemporal features can be obtained. Provides supplementary information. The core idea is to use local spatiotemporal features to query similar historical patterns in memory and output a weighted feature representation for the current timestamp. By fusing multi-scale historical features with local spatiotemporal features, it not only eliminates the negative impact of missing values ​​but also enhances the current feature embedding.

[0037] In time t , the query vector currently input into the multi-scale memory network Through local spatiotemporal characteristics Embed generation: ; in 、 is a learnable parameter, d The embedding dimension can be based on F set up.

[0038] The input of the memory module consists of multi-scale time segments: the observation data of multiple scales embedded in the runoff generation value include multiple scales of daily scale periodic data, weekly scale periodic data, monthly scale periodic data, ten-day scale periodic data and annual scale periodic data.

[0039] In a specific embodiment, the observation data at multiple scales embedded in the runoff generation value includes daily scale periodic data, weekly scale periodic data, and monthly scale periodic data. The observation data at multiple scales are specifically as follows: The data at each moment in the daily periodic data include the observation data of multiple days before that moment; The data at each moment in the weekly periodic data include the observation data of multiple days before and after the moment, centered on the day corresponding to the moment. The data at each moment in the monthly periodic data include the observation data for multiple days before and after the moment, centered on the day corresponding to the moment, in the multiple months before the moment.

[0040] Specifically, daily scale periodic data: , For daily periodic data t Time data, including t Before the time Days of data, for example, can be obtained by t Before the time The observation data of the past 24 hours are spliced ​​together.

[0041] Weekly scale periodic data: , For weekly periodic data t Time data, including t Observation data spanning multiple weeks before the moment, where The time value of the observation data included is , that is, in t Time passed Midweek, weekly storage t Time corresponds to the day before and after Observation data of is the number of samples per week, Indicates a concatenation operation. For example, 、 、 t The time corresponds to Wednesday, and the weekly period data t The data at this moment include t Observation data for Monday, Tuesday, Thursday, and Friday 2 weeks before the moment, and t Observation data for Monday, Tuesday, Thursday, and Friday of the week before the time.

[0042] Monthly scale cycle data: , For monthly periodic data t Time data, including t Observation data spanning multiple months before the moment, where The time value of the observation data included is , that is, in t Time passed Mid-month, monthly storage t Time corresponds to the day before and after A sample of is the number of samples per month. For example, 、 、 t The time corresponds to the 3rd, and the monthly scale cycle data t The data at this moment include t Observation data for February 1st, 2nd, 4th and 5th before the moment, and tObservation data for January 1, 2, 4, and 5 before the moment.

[0043] The input set Embedded as input memory vector and output memory vector : ; ; in 、 As a parameter.

[0044] The runoff generated value embedding outputs the generated value data of flood runoff at all hydrological stations corresponding to the moment of the missing value, as follows: The multi-scale memory network calculates the query vector in the embedding space through the inner product and Softmax function With each input memory vector Attention score : ; The attention score reflects the similarity between historical observations and the query, and a higher score indicates that the pattern is more relevant to the current missing value context. By outputting the memory vector The weighted sum of is generated: ; Finally, the local spatiotemporal features are fused with the global multi-scale features to output the enhanced runoff embedding: ; in, for t The generated value data of the moment runoff; 、 is the parameter, Indicates splicing; Represents the number of input memory vectors and the number of output memory vectors, wherein the input memory vectors and the output memory vectors form the memory data. , we can get the enriched runoff embedding data .

[0045] Finally, the flood forecasting model uses a hybrid model combining a graph convolutional network (GCN) and a long short-term memory (LSTM) network. The GCN extracts spatial features from the runoff data after generating value embedding, while the LSTM extracts temporal features based on these spatial features and outputs the forecast results. This creates a GCN-LSTM model for flood forecasting.

[0046] The GCN-LSTM model is a hybrid model that combines a graph convolutional network (GCN) and a long short-term memory network (LSTM). It aims to simultaneously capture spatial dependencies and temporal dynamics in data. It is widely used in scenarios requiring joint spatiotemporal modeling, such as traffic forecasting, social network analysis, and runoff prediction. Typical architectures can be divided into two modes: (1) Cascade structure: Step 1: Spatial feature extraction: The input data of each time step (such as the time series of the graph structure) first passes through the GCN layer to aggregate the information of neighboring nodes and generate a high-order spatial representation of the node.

[0047] Input: graph structure data (adjacency matrix) and node feature matrix.

[0048] Output: spatial features after graph convolution.

[0049] Step 2: Time Series Modeling: The GCN outputs of multiple time steps are input into the LSTM layer in chronological order to capture the dynamic changes in the time dimension.

[0050] Input: spatial features over a time series (e.g., GCN outputs over the past T time steps).

[0051] Output: Predictions for future time steps.

[0052] (2) Alternating structure: GCN is embedded into each time step of LSTM to form a fusion unit. At each time step, the input of LSTM is obtained by processing the current data and the hidden state of the previous time step through GCN.

[0053] This embodiment uses the first cascade structure. The input of the graph convolutional network is as follows: Establishing a network graph of the target hydrological station including a plurality of hydrological sites, wherein the network graph includes nodes consisting of the plurality of hydrological sites and edges connecting the nodes with flow relationships, wherein each node has location information and runoff data information; Based on the network graph, an adjacency matrix and a node feature matrix are established as inputs of the graph convolutional network.

[0054] The core feature of this model is to aggregate information about neighboring nodes through graph convolution operations to capture spatial dependencies between nodes. It is suitable for non-Euclidean data and can handle dynamically changing graph structures. It uses the LSTM gating mechanism (forget gate, input gate, output gate) to model long-term time dependencies and avoid the gradient vanishing problem. It is suitable for processing periodic and trend changes in time series data. The parameters of GCN and LSTM are shared between time steps, reducing model complexity. The enriched missing data values ​​and the adjacency matrix of the graph are input into the model. The GCN-LSTM framework is used as a feature extractor, and then the final prediction result is generated through the fully connected layer, such as Figure 3 shown.

[0055] Furthermore, this embodiment also provides a flood forecasting system for data-scarce areas based on multi-scale memory and deep learning, the system including a memory and a processor, the memory storing a computer program, and the processor executing any one of the above-mentioned flood forecasting methods for data-scarce areas based on multi-scale memory and deep learning when executing the computer program.

[0056] In some specific application examples, in order to evaluate the prediction effect of the method used in this embodiment, four models were compared: RNN, GRU, LSTM and GCN-LSTM (the present invention). To ensure fair comparison, all models were trained using the same configuration: ReLU activation function, mean square error loss function and Adam optimizer. The training, validation and test data sets of different models are consistent. The data set includes 60% random missing rate data, 30% mixed missing rate data, 50% mixed missing rate data, and 70% mixed missing rate data. Mixed missing includes 50% random missing and 50% block missing, where random missing means that all site data are randomly missing in time, and block missing means that all sites are missing data at a certain point in time. The input time series feature length is 14 days, and the output is the runoff forecast for the next 1 day. In the multi-scale memory network module, L and S Set to 14 and 4 respectively, Set to 4, , , are all set to 2, the embedding dimension d Set to 32.

[0057] This example uses historical runoff observations from 12 hydrological stations in Switzerland, covering a period of 14,610 days from January 1, 1981, to December 31, 2020. The data is divided into a training set and a test set in an 8:2 ratio. These two datasets are used for model training and testing, respectively, and remain completely independent.

[0058] The RNN, GRU, LSTM, and GCN-LSTM models were used to predict runoff for the next day at 12 hydrological stations in the basin. To evaluate the performance of different models, the following indicators were used: mean square error (MSE), mean absolute error (MAE), Nash efficiency coefficient (NSE), and mean relative error (MRE). The results are shown in the following table: Experimental results show that GCN-LSTM significantly outperforms other models in all indicators, indicating that it has the strongest comprehensive prediction ability. The lowest MSE indicates that the squared error between the predicted value and the true value is the smallest, the lowest MAE indicates that the absolute value of the prediction deviation is smaller, the NSE closest to 1 reflects the model's best ability to capture data changes, and the lowest MRE indicates optimal relative error control. However, the MRE value is still relatively high, which may be due to the characteristics of runoff data.

[0059] The results of running multiple sets of data show that the missing value processing method proposed in this embodiment combined with the GCN-LSTM model is the most effective, indicating that the spatial data processing capabilities of the graph neural network and the temporal processing capabilities of the long-short-term memory network combined with the memory network rich in missing data can achieve accurate flood forecasting for small watersheds.

[0060] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A flood forecasting method for data-scarce areas based on multi-scale memory and deep learning, characterized by: include: Runoff collection: Collect observation data of flood runoff from multiple hydrological stations in the target hydrological station; Missing value feature extraction: For missing values ​​in the observation data, the time correlation of the missing value is obtained based on the observation information close to the time of the missing value, and the spatial correlation of the missing value is obtained based on the observation information close to the hydrological station where the missing value is located; the single-point spatiotemporal characteristics of the missing value are obtained based on the time correlation and spatial correlation; and then the local spatiotemporal characteristics corresponding to the moment of the missing value, including the single-point spatiotemporal characteristics corresponding to all hydrological stations, are obtained; Runoff generated value embedding: The local spatiotemporal features are input into a multi-scale memory network to generate a query vector, and based on the query vector and the memory data in the multi-scale memory network, generated value data including flood runoff at all hydrological stations corresponding to the time of the missing value is output to achieve runoff embedding; wherein the memory data is obtained by inputting observation data at multiple scales into the multi-scale memory network; Flood forecasting: Based on the runoff data after the runoff generation value is embedded, flood forecasting is performed using a prediction model.

2. The flood forecasting method for data-deficient areas based on multi-scale memory and deep learning according to claim 1, characterized in that: The time-related quantity of the missing value obtained in the missing value feature extraction specifically includes: The empirical time mean is obtained by using multiple observations of the hydrological station where the missing value is located before the time corresponding to the missing value. ; Get the last time observation value based on the closest and complete observation value before the corresponding time of the missing value at the hydrological station where the missing value is located and the time interval between the two ; Accordingly, obtaining the spatial correlation of the missing value specifically includes: The empirical spatial mean is obtained by using the observation values ​​corresponding to multiple hydrological stations close to the hydrological station where the missing value is located. ; Obtain the nearest spatial observation value based on the observation value of the hydrological station that is closest to and complete to the hydrological station where the missing value is located and the spatial distance between the two .

3. The flood forecasting method for data-deficient areas based on multi-scale memory and deep learning according to claim 2, characterized in that: Experience time mean The formula is as follows: ； Empirical space mean The formula is as follows: ； The single-point spatiotemporal characteristics of the missing value As shown in the following formula: ； Where, For the i The first hydrological station t The observed value at the time; is a mask, 0 indicates that the corresponding observation is missing, and 1 indicates that the corresponding observation is complete; L Indicates taking t Before the moment L Observations are used to obtain the empirical time mean; S Indicates taking the hydrological station close to the missing value S The empirical spatial mean of the observation values ​​corresponding to the hydrological stations is obtained; is the time decay rate; is the spatial attenuation rate.

4. The flood forecasting method for data-deficient areas based on multi-scale memory and deep learning according to claim 3, characterized in that: The spatiotemporal decay rate is modeled as: ； in, , For hydrological stations i The corresponding time decay parameter controls the influence of time distance; ； in, , For hydrological stations i The corresponding spatial attenuation parameter controls the influence of spatial distance.

5. The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning according to any one of claims 1 to 4, characterized in that: The observation data of multiple scales in the runoff generation value embedding include multiple scale periodic data of daily scale, weekly scale, monthly scale, ten-day scale and annual scale.

6. The flood forecasting method for data-deficient areas based on multi-scale memory and deep learning according to claim 5, characterized in that: The observation data at multiple scales in the runoff generation value embedding include daily scale periodic data, weekly scale periodic data, and monthly scale periodic data. The observation data at multiple scales are as follows: The data at each moment in the daily periodic data include the observation data of multiple days before that moment; The data at each moment in the weekly periodic data include the observation data of multiple days before and after the corresponding day of the moment in the weeks before the moment; The data at each moment in the monthly periodic data include the observation data for multiple days before and after the moment, centered on the day corresponding to the moment, in the multiple months before the moment.

7. The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning according to any one of claims 1 to 4, characterized in that: The runoff generated value embedding outputs the generated value data of flood runoff at all hydrological stations corresponding to the moment of the missing value, as follows: The multi-scale memory network calculates the query vector through inner product and Softmax function With each input memory vector Attention score : ； The attention score reflects the similarity between historical observations and queries. A higher score indicates that the pattern is more relevant to the current missing value context. The memory response vector By outputting the memory vector The weighted sum of is generated: ； Finally, the output enhanced runoff embedding is as follows: ； in, for t Generated value data of runoff at each moment; 、 As parameters, Indicates splicing, represents the number of input memory vectors and the number of output memory vectors, wherein the input memory vectors and the output memory vectors form the memory data.

8. The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning according to any one of claims 1 to 4, characterized in that: The prediction model in flood forecasting adopts a hybrid model that combines a graph convolutional network and a long short-term memory network, where the graph convolutional network is used to extract the spatial features of runoff data after the generated values ​​are embedded, and the long short-term memory network is used to extract temporal features based on the spatial features and output the forecast results.

9. The flood forecasting method for data-deficient areas based on multi-scale memory and deep learning according to claim 8, characterized in that: The input of the graph convolutional network is as follows: Establishing a network graph of the target hydrological station including a plurality of hydrological sites, wherein the network graph includes nodes consisting of the plurality of hydrological sites and edges connecting the nodes with flow relationships, wherein each node has location information and runoff data information; Based on the network graph, an adjacency matrix and a node feature matrix are established as inputs of the graph convolutional network.

10. A flood forecasting system for data-scarce areas based on multi-scale memory and deep learning, characterized by: The system includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it executes the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning described in any one of claims 1 to 9.

Images