Flood forecasting method and system based on multi-scale memory and deep learning in data-scarce areas

Abstract

The present application belongs to the technical field of hydrological forecast, and discloses a flood forecasting method and system in a region with insufficient data based on multi-scale memory and deep learning, wherein the method comprises the following steps: collecting observation value data of flood runoff of multiple hydrological stations; for missing values, acquiring time correlation quantity according to time-close observation value information, acquiring space correlation quantity according to space-close observation value information, so as to acquire single-point space-time characteristics of the missing values; acquiring local space-time characteristics including characteristics of all hydrological stations; inputting the local space-time characteristics into a multi-scale memory network storing multiple scale observation value data to output generated value data to realize runoff embedding; and performing flood forecasting based on runoff data after runoff embedding. The present application comprehensively considers space-time relationship of hydrological stations to extract characteristics of missing values, and fuses multi-scale historical characteristics and local space-time characteristics through a multi-scale memory network structure, so that current characteristic embedding can be enhanced, thereby improving prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of hydrological forecasting technology, and more specifically, relates to a flood forecasting method and system for data-scarce areas based on multi-scale memory and deep learning. Background Technology

[0002] Missing time-series data is a common problem in flood forecasting modeling. Missing data makes it difficult to capture the evolution patterns of the entire sequence and prevents the prediction model from fully extracting valuable features from the sequence, resulting in a significant reduction in prediction accuracy. Hydrological data in flood forecasting is a type of time-series data that fits this characteristic. Based on the handling of missing values, there are two main methods for processing missing data.

[0003] Deletion method: This method directly deletes the entire feature variable with missing values. Its advantage is that it solves the problem of missing values, but its disadvantage is that the reduction in data leads to a significant loss of information, which is extremely detrimental to the prediction model's ability to mine features and evolution patterns in time series, often resulting in poor predictive performance. Interpolation method: The interpolation method does not delete any sample data; instead, it fills in the missing data using various methods, ensuring data integrity. The development of hydrological data missing value imputation techniques can be divided into three stages.

[0004] Early methods relied primarily on traditional statistical techniques, with classical techniques such as linear interpolation, Lagrange interpolation, and polynomial interpolation dominating for a long time. These methods estimate missing values ​​by establishing linear or polynomial relationships between adjacent data points, proving well-suited for scenarios with low data fluctuations and limited missing value ranges. With technological advancements, methods like 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 complex data distribution characteristics when dealing with highly heterogeneous data. The introduction of time series models has opened new avenues for missing value imputation. Time series analysis methods, represented by ARIMA and seasonality models, predict missing values ​​by mining historical evolution patterns in the data. These methods offer the advantage of ease of implementation, but their prediction accuracy may be limited when handling complex time series containing multi-scale features.

[0005] In recent years, breakthroughs in artificial intelligence technology have driven innovation in the field of hydrological data imputation. Machine learning algorithms have demonstrated powerful nonlinear modeling capabilities. For example, researchers have built an interpolation model for annual precipitation data in the Three Gorges area based on Support Vector Machines (SVM), and the Random Forest method has achieved better results than Kriging interpolation on a synthetic dataset in the Spanish region. The introduction of deep neural network models has brought about a paradigm shift. Researchers have proposed an improved recurrent neural network (RNN) architecture, successfully solving the noise interference problem in Automatic Calibration Signal (ACS) data; and combining Long Short-Term Memory (LSTM) networks with Hermitian spline interpolation (CHS) has enabled the effective recovery of high-resolution data from Vessel Monitoring Systems (VMS). These methods significantly improve the accuracy of missing value imputation by capturing long-distance dependencies in the data.

[0006] However, flood runoff data not only has temporal correlation but also spatial correlation. Spatially connected watersheds can also affect flood runoff. Existing flood prediction models mainly deal with missing values ​​from the time dimension, without fully considering the complex spatiotemporal characteristics of flood runoff data, and the prediction accuracy needs to be further improved. Summary of the Invention

[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a flood forecasting method and system for data-scarce areas based on multi-scale memory and deep learning. This method solves the problem that existing flood prediction models mainly process missing values ​​from a time dimension, failing to fully consider the complex spatiotemporal characteristics of flood runoff data, and thus requiring further improvement in prediction accuracy.

[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: Collecting observed data on flood runoff from multiple hydrological stations within the target hydrological station; Missing value feature extraction: For missing values ​​in the observation data, the temporal correlation of the missing value is obtained based on the observation information that is close to the time of the missing value, and the spatial correlation of the missing value is obtained based on the observation information that is spatially close to the hydrological station where the missing value is located; the single-point spatiotemporal features of the missing value are obtained based on the temporal and spatial correlations; and then the local spatiotemporal features corresponding to the time of the missing value, including the single-point spatiotemporal features corresponding to all hydrological stations, are obtained. Runoff generation value embedding: The local spatiotemporal features are input into a multi-scale memory network to generate a query vector, and the generated value data of flood runoff at all hydrological stations corresponding to the time of the missing value is output according to the query vector and the memory data in the multi-scale memory network to realize runoff embedding; wherein, the memory data is obtained by inputting observation data at multiple scales into the multi-scale memory network; Flood forecasting: Flood forecasting is performed using predictive models based on runoff data embedded with runoff generation values.

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

[0010] The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning provided by this invention, empirical time mean The formula is as follows: ; Mean of the experience space The formula is as follows: ; The single-point spatiotemporal characteristics of the missing value As shown in the following formula: ; In the formula, For the first i The first hydrological station t The observed value at time; The mask is set to 0 to indicate that the corresponding observation is missing, and 1 to indicate that the corresponding observation is complete. L Indicates taking t Before the moment LThe empirical time mean is obtained from each observation; S This indicates that the value should be taken from the hydrological station closest to the location of the missing value. S The empirical spatial mean was obtained from the observations corresponding to each hydrological station. The decay rate over time; This represents the spatial decay rate.

[0011] According to the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning provided by this invention, the spatiotemporal decay rate is modeled as follows: ; in, , Hydrological stations i The corresponding time decay parameter controls the intensity of the effect of time distance; ; in, , Hydrological stations i The corresponding spatial attenuation parameter controls the intensity of 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 runoff generation value is embedded in observation data at multiple scales, including daily, weekly, monthly, ten-day, and annual periodic data.

[0013] According to the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning provided by this invention, the runoff generation value embedding includes observational data at multiple scales, including daily, weekly, and monthly periodic data. The specific observational data at these multiple scales are as follows: The data at each moment in diurnal periodic data includes observation data from multiple days prior to that moment; The data at each moment in the weekly-scale periodic data includes observation data from multiple weeks prior to that moment, centered on the day corresponding to that moment; The data for each moment in the monthly-scale periodic data includes observations from the months preceding that moment, centered on the day corresponding to that moment, for several days before and after that moment.

[0014] According to the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning provided by the present invention, the runoff generation value embedding outputs the generation value data of flood runoff from all hydrological stations corresponding to the time of the missing value, as detailed below: The multi-scale memory network calculates the query vector using the inner product and the Softmax function. With each input memory vector Attention score : ; Attention score reflects the similarity between historical observations and queries; a higher score indicates a greater relevance of the pattern to the current missing value context; memory response vector. By outputting the memory vector Weighted sum generation: ; Finally, the enhanced runoff embedding is output as follows: ; in, for t Generated values ​​of runoff at different times; , For parameters, Indicates splicing, This represents the number of input memory vectors and the number of output memory vectors, where the input memory vectors and 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 the flood forecasting adopts a hybrid model combining graph convolutional networks and long short-term memory networks. The graph convolutional network is used to extract spatial features of runoff data after value embedding, and the long short-term memory network is used to extract 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: Establish a network diagram of multiple hydrological stations for the target hydrological station. The network diagram includes nodes composed of multiple hydrological stations and edges connecting nodes with flow relationships. Each node has location information and runoff data information. The adjacency matrix and node feature matrix are established based on the network graph and used as inputs to 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 performs the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning described above.

[0018] Overall, compared with the prior art, the flood forecasting method and system for data-scarce areas based on multi-scale memory and deep learning provided by this invention offer the following advantages: 1. Considering the strong spatial correlation between hydrological stations, which can provide more information for missing value processing, this paper proposes to extract missing value features by comprehensively considering the spatiotemporal relationship of hydrological stations, and to store multi-scale global historical patterns through a multi-scale memory network structure. By embedding multi-scale historical observation data into memory, supplementary information can be provided for local spatiotemporal features. By fusing multi-scale historical features with local spatiotemporal features, not only can the negative impact of missing values ​​be eliminated, but the current feature embedding can also be enhanced, thereby providing more useful information for data-scarce areas and improving prediction accuracy. 2. For missing values, the empirical time mean, the last time observation, the time interval between the last time observation and the last time observation, the empirical spatial mean, the most recent spatial observation, and the spatial distance between the most recent spatial observation and the most recent spatial observation are extracted as correlation quantities to construct single-point spatiotemporal features of the missing values, thereby forming local spatiotemporal features. This can integrate temporal and spatial information to improve the performance and reliability of the generated value enhancement embedding. The constructed exponential spatiotemporal decay rate can better and more reasonably balance the degree of influence of each correlation quantity, improving intelligence and flexibility. 3. Constructing a hybrid model combining Graph Convolutional Network (GCN) and Long Short-Term Memory Network (LSTM) for flood forecasting based on enhanced embedded runoff data. The model inputs the enriched missing data values ​​and the adjacency matrix of the graph. The GCN-LSTM framework is used as a feature extractor, and the final prediction result is generated through fully connected layers. This model can simultaneously capture the spatial dependencies and temporal dynamics of the data, which is beneficial to improving the prediction accuracy. Attached Figure Description

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

[0020] Figure 2 This is a logical schematic diagram of how the multi-scale memory network in this embodiment of the invention enriches runoff embedding using local spatiotemporal features and global historical patterns.

[0021] Figure 3 This is a schematic diagram of the GCN-LSTM model structure in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be 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 illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] Please see Figure 1This embodiment provides a flood forecasting method for data-scarce areas based on multi-scale memory and deep learning. The method includes: Runoff collection: Collecting observed data on flood runoff from multiple hydrological stations within the target hydrological station; Missing value feature extraction: For missing values ​​in the observation data, the temporal correlation of the missing value is obtained based on the observation information that is close to the time of the missing value, and the spatial correlation of the missing value is obtained based on the observation information that is spatially close to the hydrological station where the missing value is located; the single-point spatiotemporal features of the missing value are obtained based on the temporal and spatial correlations; and then the local spatiotemporal features corresponding to the time of the missing value, including the single-point spatiotemporal features corresponding to all hydrological stations, are obtained. Runoff generation value embedding: The local spatiotemporal features are input into a multi-scale memory network to generate a query vector, and the generated value data of flood runoff at all hydrological stations corresponding to the time of the missing value is output according to the query vector and the memory data in the multi-scale memory network to realize runoff embedding; wherein, the memory data is obtained by inputting observation data at multiple scales into the multi-scale memory network; Flood forecasting: Flood forecasting is performed using predictive models based on runoff data embedded with runoff generation values.

[0024] The technical problem to be solved by the embodiments of the present invention is to overcome the defect of existing flood prediction technology for watersheds with scarce data that does not consider the strong spatial correlation between hydrological stations, and to provide a flood prediction technology for watersheds with scarce data based on the spatiotemporal relationship of stations, thereby providing more useful information for areas with scarce data and improving 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 set of edges, where connecting edges are set between two stations with a flow relationship. The observation data for each station at each time step contains... F Features such as runoff, rainfall, and evapotranspiration. Use This indicates that all sites are in All feature values ​​within a time slice. Indicates time t All feature values ​​of all sites, and yes Middle i The feature value vectors of each site. Simultaneously, a mask sequence is defined. ,in ,here Indicates the first i Each site at timet The feature value is missing. If If missing, then otherwise The goal of this embodiment is to construct a model. It can produce incomplete runoff sequences and hydrological station network As input, predict the future Runoff data at each time step: .

[0026] In some specific embodiments, considering the characteristics of small watersheds such as complex terrain and rapid flow velocity, and the lack of data in data-scarce areas, traditional flood forecasting methods are difficult to apply effectively. This embodiment aims to utilize the spatial data processing capabilities of graph neural networks and the temporal processing capabilities of recurrent neural networks, combined with a memory network to enrich missing data, to achieve accurate flood forecasting for such areas. This specific embodiment consists of the following steps: First, the missing data is enriched by using a multi-scale memory network to consider the spatiotemporal correlation between hydrological stations. Second, a graph convolutional network (GCN) is used to capture the spatial correlation between node attributes, while a long short-term memory (LSTM) network is used to simulate the temporal correlation based on historical data. This combined method 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 generates the final prediction result.

[0027] First, multi-scale memory networks are used to enrich and enhance runoff data: To enrich the input data by leveraging the complex spatiotemporal relationships among hydrological stations, a memory network was employed. The main idea of ​​this network is to learn from historical memory components that preserve long-range, multi-scale patterns, namely daily, weekly, and monthly cycle dependencies. The scale range depends on the data characteristics. Specifically, local spatiotemporal features are extracted as key to querying the memory components. These features include short-term temporal correlation features obtained from the station containing the missing runoff value, and spatial correlation features obtained from neighboring stations. Weighted historical long-term patterns—i.e., the memory data—are combined with local statistical features to mitigate the negative impact of missing values. Finally, the local-global features are output as a rich flow embedding.

[0028] refer to Figure 2 First, a feature extraction module can be constructed to extract features of missing values ​​from the original observation data, and then extract spatiotemporal features using the contextual information of the time series observations. 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 from multiple observations taken by the hydrological station where the missing value is located before the time corresponding to the missing value. The last time observation is obtained from the hydrological station containing the missing value and the closest complete observation before the time corresponding to the missing value. and the time interval between them ; to obtain the time-related quantity of the missing value.

[0029] Average of experience time Missing values Experience time mean through time t Previous L Calculated from historical observations, reflecting recent conditions and providing insights into... The context information. The specific formula is as follows: ; Indicates time step Time Station Whether the observed data is missing (1 for complete, 0 for missing) (using (Used as a mask to create missing values ​​in the complete dataset).

[0030] Last time observation: Based on the assumption that any missing value will inherit some information from the previous non-missing observation. That is, temporally adjacent values ​​are close to the current missing value. (This assumes missing values...) Information can be obtained from its last valid observation. Inheritance, the time interval between the two is (For example, taking the number of steps at intervals).

[0031] Second, the empirical spatial mean is obtained by using observations from multiple hydrological stations near the hydrological station where the missing value is located. The nearest spatial observation is obtained from the observations of the hydrological station that is closest to and has a complete record of the missing value. and the spatial distance between them ; to obtain the spatial correlation of the missing value.

[0032] Mean of the experience space For each missing value Utilizing hydrological stations i Nearby S Each sample can be taken from the station. i Recent S The observations from each station are used to form a sample, which is then used to construct its empirical spatial mean, reflecting the local runoff status. The specific formula is as follows: ; Indicates that the neighboring site s is at time step t The mask value.

[0033] Recent spatial observations: Typically, the state of a graph node is similar to that of its neighbors, and neighboring stations share similar runoff conditions. Definition For nodes i At time step t The spatial distance to the nearest and most complete spatial observation (such as the geographically closest node) is . (e.g., Euclidean distance).

[0034] Spatiotemporal decay rate: Based on spatiotemporal distance, a trainable time decay rate is designed. and spatial decay rate This is used to dynamically weigh the contribution of the last observation to the mean. Generally, when and When they are younger, they are more inclined to believe and As spatial / temporal distances increase, spatial / temporal averages become more representative. Based on this assumption, the spatiotemporal decay rate... The model is as follows: ; in, , Hydrological stations i The corresponding time decay parameter controls the intensity of the effect of time distance; ; in, , Hydrological stations i The corresponding spatial attenuation parameters control the intensity of the influence of spatial distance. Both time and spatial attenuation parameters can be determined through training. An exponentially oriented negative rectifier is chosen to achieve the desired attenuation rate. and It decreases monotonically within the range of 0 to 1.

[0035] Considering trainable decay, the proposed model incorporates spatial / temporal estimation to define missing values. Single-point spatiotemporal characteristics As shown in the following formula: ; In the formula, For the first i The first hydrological station t The observed value at time; The mask is set to 0 to indicate that the corresponding observation is missing, and 1 to indicate that the corresponding observation is complete. LIndicates taking t Before the moment L The empirical time mean is obtained from each observation; S This indicates that the value should be taken from the hydrological station closest to the location of the missing value. S The empirical spatial mean was obtained from the observations corresponding to each hydrological station. The decay rate over time; This represents the spatial decay rate. Therefore, for missing values... This allows us to obtain the local spatiotemporal characteristics at the corresponding moment. .

[0036] Then, a multi-scale memory network structure is constructed, with global historical patterns playing a crucial role in establishing rich runoff embedding data. By embedding multi-scale (daily, weekly, monthly) historical observation data into memory, local spatiotemporal features can be obtained. This provides supplementary information. Its core idea is to leverage 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 current input query vector for the multi-scale memory network Through local spatiotemporal features Embedding generation: ; in , For learnable parameters, d For embedding dimensions, it can be based on F set up.

[0038] The input to the memory module consists of multi-scale time segments: the runoff generation value embedding includes observation data at multiple scales, including daily, weekly, monthly, decadal, and annual periodic data.

[0039] In one specific embodiment, the runoff generation value embedding includes observational data at multiple scales, including daily, weekly, and monthly periodic data. The specific observational data at multiple scales are as follows: The data at each moment in diurnal periodic data includes observation data from multiple days prior to that moment; The data at each moment in the weekly-scale periodic data includes observation data from multiple weeks prior to that moment, centered on the day corresponding to that moment; The data for each moment in the monthly-scale periodic data includes observations from the months preceding that moment, centered on the day corresponding to that moment, for several days before and after that moment.

[0040] Specifically, daily-scale periodic data: , For daily-scale periodic data t Time-based data, including t Time before Data from the day, for example, can be obtained from t Time before The data is created by stitching together observational data from each day.

[0041] Weekly periodic data: , For periodic data on a weekly scale t Time-based data, including t Observational data spanning multiple weeks prior to the specified time, among which The included observation data has a time value of That is, in t Time passed Midweek, weekly storage t The time corresponds to the days before and after. The observed data, This refers to the number of samples per week. This indicates a concatenation operation. For example, in... , , t When the time corresponds to Wednesday, in the weekly periodic data t The data at any given time includes t Observational data from Mondays, Tuesdays, Thursdays, and Fridays two weeks prior to the time point, and t Observational data from Monday, Tuesday, Thursday, and Friday of the week preceding the time point.

[0042] Monthly periodic data: , For monthly periodic data t Time-based data, including t Observational data spanning multiple months prior to the specified time, including The included observation data has a time value of That is, in t Time passed Mid-month, monthly storage t The time corresponds to the days before and after. The sample, This refers to the number of samples in a single month. For example, in... , , t The time corresponds to the 3rd, in the monthly-scale periodic data. t The data at any given time includes t Observational data for February 1st, 2nd, 4th, and 5th prior to the specified time, and tThe observation data for January 1st, 2nd, 4th and 5th prior to the specified time.

[0043] input set Embedded as input memory vector and output memory vector : ; ; in , For parameters.

[0044] The runoff generation value embedding outputs the flood runoff generation value data for all hydrological stations corresponding to the time of the missing value, as detailed below: The multi-scale memory network computes the query vector in the embedding space using the inner product and the Softmax function. With each input memory vector Attention score : ; Attention scores reflect the similarity between historical observations and queries; higher scores indicate a greater relevance of the pattern to the current missing value context. (Memory response vector) By outputting the memory vector Weighted sum generation: ; Finally, by fusing local spatiotemporal features with global multi-scale features, an enhanced runoff embedding is output: ; in, for t Generated values ​​of runoff at different times; , For parameters, Indicates splicing; This represents the number of input memory vectors and the number of output memory vectors, where the input memory vectors and output memory vectors form the memory data. For the input... This allows for the acquisition of richer runoff embedding data. .

[0045] Finally, the flood forecasting model employs a hybrid model combining a Graph Convolutional Network (GCN) and a Long Short-Term Memory (LSTM) network. The GCN is used to extract spatial features from the generated value-embedded runoff data, while the LSTM is used to extract temporal features based on these spatial features and output the forecast results. In other words, a GCN-LSTM model is constructed for flood forecasting.

[0046] The GCN-LSTM model is a hybrid model combining Graph Convolutional Networks (GCN) and Long Short-Term Memory Networks (LSTM) 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 categorized into two modes: (1) Cascaded structure: Step 1: Spatial Feature Extraction The input data at each time step (time series as shown in the figure) first passes through the GCN layer, which aggregates information from neighboring nodes to generate a high-order spatial representation of the nodes.

[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 outputs of GCN at multiple time steps are fed into the LSTM layer in chronological order to capture dynamic changes in the time dimension.

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

[0051] Output: Predictions for future time steps.

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

[0053] This embodiment uses the first type of cascaded structure. The input to the graph convolutional network is as follows: Establish a network diagram of multiple hydrological stations for the target hydrological station. The network diagram includes nodes composed of multiple hydrological stations and edges connecting nodes with flow relationships. Each node has location information and runoff data information. The adjacency matrix and node feature matrix are established based on the network graph and used as inputs to the graph convolutional network.

[0054] The core feature of this model is that it aggregates information from 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 utilizes the gating mechanism of LSTM (forget gate, input gate, output gate) to model long-term time dependencies, avoiding the gradient vanishing problem, making it suitable for handling periodic and trend-based 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 graph's adjacency matrix are input into the model. The GCN-LSTM framework is used as a feature extractor, and then the final prediction result is generated through fully connected layers, such as... Figure 3 As 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 includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning described above.

[0056] In specific application examples, to evaluate the predictive performance of the methods used in this embodiment, four models were compared: RNN, GRU, LSTM, and GCN-LSTM (this invention). To ensure a fair comparison, all models were trained using the same configuration: ReLU activation function, mean squared error loss function, and Adam optimizer. The training, validation, and testing datasets for the different models were consistent. The datasets included data with 60% random missing rate, 30% mixed missing rate, 50% mixed missing rate, and 70% mixed missing rate. Mixed missing rates included 50% random missing and 50% block missing, where random missing refers to all station data being randomly missing in time, and block missing refers to all stations having data missing at a certain point in time. The input time series feature length was 14 days, and the output was the runoff prediction for the next day. In the multi-scale memory network module, L and S Set them to 14 and 4 respectively. Set to 4, , , All are set to 2, embedding dimension d Set it to 32.

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

[0058] Runoff forecasts for the next day were performed at 12 hydrological stations within the watershed using RNN, GRU, LSTM, and GCN-LSTM models. To evaluate the performance of different models, the following metrics were used: mean squared error (MSE), mean absolute error (MAE), Nash efficiency coefficient (NSE), and mean relative error (MRE). The results are shown in the table below. 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 the smallest squared error between the predicted and actual values, the lowest MAE indicates a smaller absolute value of the prediction bias, the NSE being closest to 1 reflects the model's best ability to capture data changes, and the lowest MRE indicates the best relative error control. However, the MRE value is still relatively high, which may be due to the characteristics of runoff data.

[0059] The results from multiple sets of data show that the missing value handling method proposed in this embodiment, combined with the GCN-LSTM model, has the most significant effect. This indicates that by combining the spatial data processing capability of graph neural networks and the temporal processing capability of long short-term memory networks with memory networks containing abundant missing data, accurate flood forecasting for small watersheds can be achieved.

[0060] Those skilled in the art will readily understand that the above description is merely 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 within 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 in that, include: Runoff collection: Collecting observed data on flood runoff from multiple hydrological stations within the target hydrological station; Missing value feature extraction: For missing values ​​in the observation data, the temporal correlation of the missing value is obtained based on the observation information that is close to the time of the missing value, and the spatial correlation of the missing value is obtained based on the observation information that is spatially close to the hydrological station where the missing value is located; the single-point spatiotemporal features of the missing value are obtained based on the temporal and spatial correlations; and then the local spatiotemporal features corresponding to the time of the missing value, including the single-point spatiotemporal features corresponding to all hydrological stations, are obtained. Runoff generation value embedding: The local spatiotemporal features are input into a multi-scale memory network to generate a query vector, and the generated value data of flood runoff at all hydrological stations corresponding to the time of the missing value is output according to the query vector and the memory data in the multi-scale memory network to achieve runoff embedding; wherein, the memory data is obtained by inputting observation data at multiple scales into the multi-scale memory network, embedding the observation data at multiple scales into the multi-scale memory network to generate input memory vectors and output memory vectors, and the input memory vectors and output memory vectors form the memory data; Flood forecasting: Flood forecasting is performed using a predictive model based on runoff data embedded with runoff generation values; The runoff generation value embedding outputs the flood runoff generation value data for all hydrological stations corresponding to the time of the missing value, as detailed below: The multi-scale memory network calculates the query vector using the inner product and the Softmax function. With each input memory vector Attention score : ； Attention score reflects the similarity between historical observations and queries; a higher score indicates a greater relevance of the pattern to the current missing value context; memory response vector. By outputting the memory vector Weighted sum generation: ； Finally, the enhanced runoff embedding is output as follows: ； in, for t Generated values ​​of runoff at different times; , For parameters, Indicates splicing, This represents the number of input memory vectors and the number of output memory vectors.

2. The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning as described in claim 1, characterized in that, The time-related quantities for obtaining the missing values ​​in the missing value feature extraction specifically include: The empirical time mean was obtained from multiple observations taken at the hydrological station where the missing value was located before the time corresponding to the missing value. The last time observation is obtained from the hydrological station containing the missing value and the closest complete observation before the time corresponding to the missing value. and the time interval between them ; Accordingly, obtaining the spatial correlation of the missing values ​​specifically includes: The empirical spatial mean was obtained by using observations from multiple hydrological stations near the station where the missing value was located. The nearest spatial observation is obtained from the observations of the hydrological station that is closest to and has a complete record of the missing value. and the spatial distance between them .

3. The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning as described in claim 2, characterized in that, Average time of experience The formula is as follows: ； Mean of the experience space The formula is as follows: ； The single-point spatiotemporal characteristics of the missing value As shown in the following formula: ； In the formula, For the first i The first hydrological station t The observed value at time; The mask is set to 0 to indicate that the corresponding observation is missing, and 1 to indicate that the corresponding observation is complete. L Indicates taking t Before the moment L The empirical time mean is obtained from each observation; S This indicates that the value should be taken from the hydrological station closest to the location of the missing value. S The empirical spatial mean was obtained from the observations corresponding to each hydrological station. The decay rate over time; This represents the spatial decay rate.

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

5. The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning as described in any one of claims 1-4, characterized in that, The runoff generation values ​​are embedded in observational data at multiple scales, including daily, weekly, monthly, decadal, and annual periodic data.

6. The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning as described in claim 5, characterized in that, The runoff generation values ​​are embedded with observational data at multiple scales, including daily, weekly, and monthly periodic data. The specific observational data at these multiple scales are as follows: The data at each moment in diurnal periodic data includes observation data from multiple days prior to that moment; The data at each moment in the weekly-scale periodic data includes observation data from multiple weeks prior to that moment, centered on the day corresponding to that moment; The data for each moment in the monthly-scale periodic data includes observations from the months preceding that moment, centered on the day corresponding to that moment, for several days before and after that moment.

7. The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning as described in any one of claims 1-4, characterized in that, The prediction model in flood forecasting adopts a hybrid model combining graph convolutional networks and long short-term memory networks. The graph convolutional network is used to extract spatial features of runoff data after value embedding, and the long short-term memory network is used to extract temporal features based on the spatial features and output forecast results.

8. The flood forecasting method for data-scarce areas based on multi-scale memory and deep learning as described in claim 7, characterized in that, The input to the graph convolutional network is as follows: Establish a network diagram of multiple hydrological stations for the target hydrological station. The network diagram includes nodes composed of multiple hydrological stations and edges connecting nodes with flow relationships. Each node has location information and runoff data information. The adjacency matrix and node feature matrix are established based on the network graph and used as inputs to the graph convolutional network.

9. A flood forecasting system for data-scarce areas based on multi-scale memory and deep learning, characterized in that, The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it performs the flood forecasting method for data-scarce areas based on multi-scale memory and deep learning as described in any one of claims 1-8.

Images