A plain river network hydrological model reasoning method and system based on causal decoupling

Abstract

This invention discloses a method and system for inference of hydrological models in plain river networks based on causal decoupling. The method includes: collecting multi-source hydrological time series data; standardizing and preprocessing the data, dividing it into training / validation / test sets according to time, and initially screening periodic features using a sliding window method; mining statistical causal relationships using the PCMCI algorithm, incorporating hydrological physical mechanism constraints to filter out spurious associations; constructing a hydrological causal graph with time delay characteristics, and outputting causal and time-delay priors; realizing causal decoupling representation learning through a causal constraint encoder, mapping it to decoupling potential states; inputting the decoupling states into a time series prediction network, and combining a recursive multi-step prediction strategy to infer future hydrological states; and outputting prediction results such as water level and flow rate. This invention achieves physical interpretability of hydrological model inference, effectively improves the generalization ability and medium-to-long-term time series prediction accuracy under extreme hydrological events, and can provide precise technical support for flood control, drainage, and water resource allocation in plain river network basins.

Description

Technical Field

[0001] This invention relates to the field of intelligent hydrological computing technology, specifically to a reasoning method and system for a plain river network hydrological model based on causal decoupling. Background Technology

[0002] Plain river network areas are characterized by gentle terrain, dense river systems, and strong artificial regulation. Their hydrological processes are typically driven by multiple factors, including precipitation, upstream inflow, tidal backwater, gate scheduling, and pumping station operation. Significant time-lag effects, nonlinear coupling relationships, and multi-scale periodic variations exist among these driving factors, resulting in highly complex hydrological processes in plain river networks.

[0003] Existing hydrological forecasting methods are mainly divided into two categories: The first is traditional physical mechanism-driven hydrological models, including hydrodynamic models and conceptual hydrological models. These models simulate hydrological processes by establishing watershed water balance or hydrodynamic equations, but they typically require extensive calibration of river cross-section parameters, roughness parameters, and hydraulic engineering scheduling rules. The model construction and parameter calibration process is complex and difficult to adapt to plain river network areas with scarce data or frequent human intervention. The second is data-driven models. In recent years, time-series forecasting models based on deep learning, such as recurrent neural networks, convolutional neural networks, and Transformers, have been widely applied in the field of hydrological forecasting. These methods can learn the nonlinear relationships between variables through historical data and achieve good results in short-term forecasting tasks. However, these models usually rely solely on statistical correlation for modeling, making it difficult to identify true causal driving relationships. The internal mechanisms of the models lack physical interpretability and are prone to large prediction errors in extreme hydrological events or long-term forecasting scenarios.

[0004] In recent years, causal inference and decoupled representation learning techniques have been gradually introduced into the field of hydrological time series analysis, providing new ideas for solving the above problems. Reference 1, "Identification of Causal Relationships of Watershed Hydrometeorological Elements Based on PCMCI" (Journal of Hydraulic Engineering, 2023, Vol. 54, No. 6, pp. 721-732), proposes a method for identifying causal relationships of hydrological elements based on the PCMCI algorithm, verifying the applicability of the causal discovery method. However, it only completes factor screening and does not integrate causal constraints into deep learning feature learning, making it difficult to solve the feature entanglement of multi-source factors. Reference 2, "Deep learning model fordrought prediction based on large-scale spatial causal network in the Yangtze River Basin" (Journal of Hydrology, 2025, Vol. 635, p. 132808), combines PCMCI causal discovery with spatiotemporal graph convolutional networks to construct a drought prediction model for the Yangtze River Basin. It verifies the role of causal constraints in improving hydrological prediction performance. However, this method does not incorporate causal decoupling representation learning, failing to isolate the independent influence of different driving factors on hydrological processes, and the model's interpretability still has significant room for improvement. Reference 3, "Disentangled Generative Causal Representation Learning" (International Conference on Learning Representations, 2021), proposes the DEAR causal decoupling representation model, breaking through the limitations of traditional independent prior assumptions. However, it does not integrate with the physical mechanisms of hydrological processes in plain river networks, making it unsuitable for directly adapting to engineering scenarios of hydrological time-series prediction.

[0005] In summary, existing research still has three shortcomings: First, the causal discovery process lacks physical mechanism constraints, making it difficult to eliminate spurious causal relationships in complex scenarios such as tidal backwater and gate-pump linkage; second, the integration of causal decoupling representation learning and hydrological model inference is insufficient, and the feature entanglement problem caused by multi-source factor coupling remains unresolved, making it difficult to balance accuracy and interpretability; third, it does not fully adapt to the multi-period characteristics and time lag effects of hydrological processes, and the accuracy and stability of long-term multi-step prediction need to be improved. Summary of the Invention

[0006] The technical problem to be solved by the present invention is as follows: To solve the above problem, the present invention adopts the following technical solution:

[0007] First, this invention proposes a reasoning method for a plain river network hydrological model based on causal decoupling, comprising the following steps:

[0008] S1. Collect multi-source hydrological time series data of plain river network, including precipitation, upstream water inflow, gate gravity drainage, pumping station drainage, and real-time tidal data;

[0009] S2. Standardize and preprocess the multi-source hydrological time series, divide it into training set, validation set and test set, use the sliding window method to initially screen potential periodic features, adapt to the time series modeling requirements, and determine the optimal historical window length.

[0010] S3. Based on the preprocessed hydrological time series and the optimal historical window length, conduct temporal causal relationship identification, mine the statistical causal relationship and corresponding time delay characteristics between hydrological elements, incorporate the constraints of the hydrological physical mechanism of the plain river network, including tide-water level time delay and precipitation-runoff response, filter out false associations that do not conform to hydrological laws, and obtain effective causal relationships with physical rationality.

[0011] S4. Based on the results of temporal causal relationship identification, construct a hydrological causal diagram of the plain river network that includes the influence relationship and time lag characteristics, and output the causal relationship and time lag prior as the core constraint basis for subsequent causal decoupling.

[0012] S5. Construct a causal constraint encoder based on the hydrological causal graph, embed causal relationships and time-delay priors into the structure mask and parent node constraints of the encoder, and map the preprocessed hydrological data into decoupled potential states through the encoder to achieve causal decoupling in the hidden space.

[0013] S6. Conduct temporal prediction reasoning, input the decoupled potential state sequence into the spatiotemporal graph convolutional network, capture the temporal dependency and periodic characteristics of the hydrological process, and infer the future hydrological state through a recursive multi-step prediction strategy.

[0014] S7. Output the predicted results of future water level and flow in the plain river network.

[0015] Preferably, the standardization preprocessing in step S2 employs both Min-Max normalization and Z-score normalization. Min-Max normalization linearly maps the original hydrological data to the [0,1] interval, eliminating the influence of dimensions, and is specifically calculated using the following formula:

[0016] ,

[0017] in, For the first The original values ​​of the hydrological variables at time t; The value is the normalized value; , The first The maximum and minimum values ​​of a hydrological variable throughout the entire time series;

[0018] Z-score standardization is used to transform the raw hydrological data into a standard normal distribution with a mean of 0 and a standard deviation of 1, specifically calculated using the following formula:

[0019] ,

[0020] in, These are the standardized values; For the first The mean of the time series of several hydrological variables; For the first The standard deviation of the time series of hydrological variables; the training set, validation set and test set are divided in a 7:1.5:1.5 ratio according to time order; the sliding window method sets the window scale to cover 1d, 7d, 15d, 30d, 90d and 180d, and identifies potential periodic characteristics by calculating the autocorrelation coefficient and partial correlation coefficient of the sequence, and determines the optimal historical window length for time series modeling.

[0021] Preferably, in step S3, the partial correlation coefficient is used to measure the conditional independence of two hydrological variables under a given set of conditions. The partial correlation coefficient is specifically calculated using the following formula: ,

[0022] in, Let X be the partial correlation coefficient between variables X and Y under condition set Z; To remove variables from condition set Z The set after; when When X and Y are conditionally independent, accept the null hypothesis. Let α be the inverse function of the standard normal distribution, and α be the significance level, which is set to 0.05.

[0023] The MCI criterion is used to calculate the strength of causal effects, specifically using the following formula:

[0024] ,

[0025] in, Let variable X be a time delay of Y The estimated causal effect values; for The set of parent nodes contains the lagged terms of Y itself and the lagged terms of other potential driving factors; when When a causal relationship passes the significance test and meets the physical mechanism constraints, it is determined to be a valid causal relationship.

[0026] Preferably, the hydrological causal graph of the plain river network in step S4 uses precipitation, upstream water inflow, gate gravity drainage, pumping station drainage, real-time tides, river water level, and flow as nodes, the effective causal relationships obtained in step S3 as directed edges, and the corresponding time delay characteristics as edge weights to construct a directed acyclic graph structure; the causal relationships and time delay priors include a binary adjacency matrix of causal relationships, the parent node dependencies of each node, and time delay parameters.

[0027] Preferably, in step S5, a structural mask is constructed based on the binary adjacency matrix of the plain river network hydrological causal graph to constrain the neuron connections of the encoder, specifically calculated using the following formula:

[0028] ,

[0029] in, The structure mask weight matrix of the encoder; Given a binary adjacency matrix of a hydrological cause-effect graph, if the nodes... It is a node The parent node, then Otherwise, it is 0; This is the original weight matrix of the encoder; For the Hadamardi, ensure that information is transmitted only between nodes with causal relationships;

[0030] The feature propagation process of each layer of the causal constraint encoder is calculated using the following formula:

[0031] ,

[0032] in, Let L be the network weights of the encoder's layer l. It is a non-linear activation function. This is the feature output of the encoder's l-th layer;

[0033] Based on a structural causal model, the preprocessed hydrological data is compared with historical time series windows. Mapping to decoupled potential states The specific calculation is as follows:

[0034] ,

[0035] in, for The decoupled latent state vector at time step; This is the weighted adjacency matrix of the hydrological causal graph, corresponding to the intensity of causal effects; It is the identity matrix; , It is an element-wise nonlinear invertible transformation function; For exogenous noise vectors that follow a standard normal distribution, The hidden layer neurons of the causal constraint encoder strictly follow the directed acyclic structure of the hydrological causal graph, allowing only information transmission from parent nodes to child nodes and prohibiting reverse information flow.

[0036] Preferably, the spatiotemporal graph convolutional network in step S6 receives the decoupling potential state output in step S5. As node features, and using the hydrological causal graph constructed in step S4 as spatial topological constraints, the spatiotemporal coupling characteristics between hydrological elements are captured; the model training uses mean absolute error as the core loss function, as shown in the following formula:

[0037] ,

[0038] Where N is the number of training samples; To predict the number of cross sections; This represents the predicted water level / flow rate at the k-th cross-section of the n-th sample. The corresponding measured value; the recursive multi-step prediction strategy uses the single-step prediction result of the time series prediction network as the input feature for the next prediction, and simultaneously integrates the historical time series data corresponding to the time delay characteristics determined in step S3, and iteratively executes to obtain the hydrological state sequence within the preset prediction period. The maximum prediction period is set to 1d~30d, and the single-step prediction step size is consistent with the time resolution of the original hydrological data.

[0039] Preferably, the accuracy evaluation indicators of the prediction results in step S7 include root mean square error, mean absolute error, and coefficient of determination; the root mean square error is specifically calculated using the following formula:

[0040] ,

[0041] The coefficient of determination is calculated using the following formula:

[0042] ,

[0043] in, The result is the mean of all measured values; the output result synchronously includes the core driving causal factors and contribution of the corresponding prediction result, realizing the interpretability of the hydrological model reasoning process.

[0044] Meanwhile, this invention proposes a reasoning system for a plain river network hydrological model based on causal decoupling, comprising:

[0045] The data acquisition module is configured to perform the following process: acquire multi-source hydrological time series data of the plain river network;

[0046] The data preprocessing and periodic feature screening module is configured to perform the following processes: standardize the multi-source hydrological time series, divide it into training set, validation set and test set, use the sliding window method to screen potential periodic features, and determine the optimal historical window length.

[0047] The temporal causality discovery module is configured to perform the following process: based on the preprocessed hydrological time series and the optimal historical window length, it conducts temporal causality identification, mines the statistical causal relationships and corresponding time lag characteristics between hydrological elements, incorporates the constraints of the hydrological physical mechanism of the plain river network, filters out false associations that do not conform to hydrological laws, and obtains effective causal relationships with physical rationality.

[0048] The hydrological cause-effect graph construction and prior output module is configured to perform the following process: based on the results of temporal causal relationship identification, construct a hydrological cause-effect graph of the plain river network that includes the influence relationship and time lag characteristics, and output the causal relationship and time lag prior;

[0049] The causal decoupling representation learning module is configured to perform the following process: construct a causal constraint encoder based on the hydrological causal graph, embed causal relationships and time-delay priors into the structure mask and parent node constraints of the encoder, and map the preprocessed hydrological data into decoupled latent states through the encoder to achieve causal decoupling in the hidden space;

[0050] The temporal prediction inference module is configured to perform the following process: conduct temporal prediction inference, input the decoupled potential state sequence into the spatiotemporal graph convolutional network, capture the temporal dependencies and periodic characteristics of the hydrological process, and infer the future hydrological state through a recursive multi-step prediction strategy.

[0051] The results output module is configured to perform the following process: output the predicted results of future water levels and flow rates in the plain river network.

[0052] Furthermore, the present invention proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the control method as described in the present invention.

[0053] Meanwhile, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed, it implements the steps of the method described in the present invention.

[0054] The present invention, by adopting the above technical solution, has the following beneficial effects:

[0055] 1. This invention filters out false causal relationships by introducing hydrophysical mechanism constraints and combines causal decoupling representation learning, so that the output of each layer of the model can be traced back to specific driving factors, thus breaking through the limitations of traditional black box models.

[0056] 2. This invention effectively eliminates false associations, ensures the physical rationality of causal relationships, and significantly improves the generalization ability of model inference under extreme hydrological events. The PCMCI algorithm works synergistically with physical constraints such as tide-water level time lag and precipitation-runoff response to eliminate statistical false positive associations, enabling the model to maintain stable predictive performance even in scenarios where extreme rainstorms or typhoon high tides have not been historically observed.

[0057] 3. This invention achieves causal decoupling representation of multi-source coupled hydrological time-series data, significantly improving the model's ability to represent the intrinsic driving mechanisms of hydrological processes. The causal constraint encoder maps coupling factors such as precipitation, tides, and gate / pump scheduling into mutually independent latent states, eliminating feature entanglement and enabling the model to accurately identify the independent contributions of each driving factor.

[0058] 4. This invention significantly improves the inference and prediction accuracy of core hydrological indicators such as water level and flow in plain river networks. Through joint optimization of a recursive multi-step prediction strategy and a spatiotemporal graph convolutional network, the prediction error over long lead times is reduced compared to traditional deep learning methods, providing more reliable decision support for flood control and drainage scheduling. Attached Figure Description

[0059] Figure 1 This invention relates to a schematic diagram of a reasoning method for a plain river network hydrological model based on causal decoupling.

[0060] Figure 2 This is a schematic diagram of the overall architecture of the plain river network hydrological model based on causal decoupling involved in this invention. Detailed Implementation

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0062] Example 1: This example is a specific implementation of a reasoning method for a plain river network hydrological model based on causal decoupling, which is divided into the following... Figure 1 The 7 steps shown correspond to Figure 2 The overall architecture of the model shown is as follows:

[0063] S1: Data Acquisition

[0064] Multi-source hydrological time-series data were collected from the target plain river network area, including time-period average precipitation data, upstream boundary inflow / water level data, gate gravity drainage flow data, pumping station drainage flow data, real-time tidal level data at the river mouth, and measured water level and flow data at river control sections. Figure 2 The multi-source hydrological time series input in the input layer of the architecture diagram.

[0065] S2: Data Preprocessing and Initial Screening of Periodic Features

[0066] The collected multi-source hydrological time series were standardized and preprocessed. Outliers were removed using the 3σ criterion, and missing values ​​were filled in using linear interpolation. The influence of data units was eliminated by Min-Max normalization using formula (1) or Z-score standardization using formula (2).

[0067]

[0068]

[0069] in, For the first The original values ​​of the hydrological variables at time t; The value is the normalized value; , The first The maximum and minimum values ​​of a hydrological variable throughout the entire time series;

[0070] These are the standardized values; For the first The mean of the time series of several hydrological variables; For the first Standard deviation of time series of hydrological variables;

[0071] The training, validation, and test sets were divided into training, validation, and test sets in a 7:1.5:1.5 ratio according to time sequence. A multi-scale sliding window method was adopted, with window scales covering 1d, 7d, 15d, 30d, 90d, and 180d. The autocorrelation and partial correlation coefficients of the hydrological time series were calculated under different window scales. The potential periodic characteristics of intraday fluctuations, inter-day variations, semi-monthly tides, monthly variations, and seasonal variations in the hydrological series were identified, and the optimal historical window length for time series modeling was determined, corresponding to the data preprocessing module in the architecture diagram of Figure 2.

[0072] S3: Discovery of Temporal Causality

[0073] Based on the optimal historical window length determined in step S2, time-series causal discovery is carried out: the PCMCI algorithm is used to mine statistical causal relationships in multi-source hydrological time series, and the conditional independence test is completed through formula (3):

[0074]

[0075] in, Let X be the partial correlation coefficient between variables X and Y under condition set Z; To remove variables from condition set Z The set after; when When X and Y are conditionally independent, accept the null hypothesis. Let α be the inverse function of the standard normal distribution, and α be the significance level, which is set to 0.05.

[0076] The intensity of causal effects among various hydrological elements is calculated using formula (4):

[0077]

[0078] in, Let variable X be a time delay of Y The estimated causal effect values; for The set of parent nodes contains the lagged terms of Y itself and the lagged terms of other potential driving factors; when When a causal relationship passes the significance test and meets the physical mechanism constraints, it is determined to be a valid causal relationship.

[0079] Simultaneously, it incorporates constraints from the hydrological and physical mechanisms of the plain river network, including tidal-water level time lag constraints, precipitation-runoff response law constraints, gate and pump scheduling-flow change causal constraints, and upstream and downstream water balance constraints. The statistical causal relationships output by the PCMCI algorithm are filtered to eliminate false associations that do not conform to the hydrological and physical mechanisms, ultimately obtaining physically reasonable causal relationships between hydrological elements and their corresponding time lag characteristics, corresponding to the time-series causal discovery module in the architecture diagram of Figure 2.

[0080] S4: Hydrological Cause-and-Effect Graph Construction and Prior Output:

[0081] Based on the causal relationships and time-delay characteristics of hydrological elements obtained in step S3, a directed acyclic plain river network hydrological causal graph (SCM / DAG) is constructed, using each hydrological element—precipitation, upstream inflow, gate gravity drainage, pumping station drainage, real-time tides, river level, and flow rate—as nodes, causal driving relationships as directed edges, and corresponding time-delay characteristics as edge weights. From the hydrological causal graph, a binary adjacency matrix of causal relationships, parent node dependencies of each node, and time-delay parameters are extracted. The causal relationships and time-delay priors are output as the core constraints for subsequent causal decoupling representation learning. Figure 2 The module for constructing a hydrological cause-effect graph in the architecture diagram.

[0082] S5: Causal Decoupling Representation Learning

[0083] Causal decoupling representation learning is performed to construct a causal constraint encoder. The causal relationship output from step S4 and the time-delay prior are embedded into the structure mask and parent node constraints of the encoder. The structure mask corresponds one-to-one with the adjacency matrix of the hydrological causal graph, as shown in formula (5).

[0084]

[0085] in, The structure mask weight matrix of the encoder; Given a binary adjacency matrix of a hydrological cause-effect graph, if the nodes... It is a node The parent node, then Otherwise, it is 0; This is the original weight matrix of the encoder; For the Hadamardi, ensure that information is transmitted only between nodes with causal relationships;

[0086] In this formula, a structural mask is used to hard-constrain the encoder weights, ensuring that only nodes with causal relationships have neuronal connections and information transmission. In this embodiment, the encoder adopts a 3-layer fully connected network structure, and the unidirectional feature propagation process of each layer satisfies formula (6):

[0087]

[0088] in, Let L be the network weights of the encoder's layer l. It is a non-linear activation function. This is the feature output of the encoder's l-th layer;

[0089] The encoder input consists of preprocessed hydrological time-series data and historical time-series windows. The nonlinear mapping of the structural causal model is achieved through formula (7):

[0090]

[0091] in, for The decoupled latent state vector at time step; This is the weighted adjacency matrix of the hydrological causal graph, corresponding to the intensity of causal effects; It is the identity matrix; , It is an element-wise nonlinear invertible transformation function; For exogenous noise vectors that follow a standard normal distribution, The hidden layer neurons of the causal constraint encoder strictly follow the directed acyclic structure of the hydrological causal graph, allowing only information transmission from parent nodes to child nodes and prohibiting reverse information flow.

[0092] High-dimensional coupled hydrological data are mapped to mutually decoupled latent states in the hidden space. This achieves causal decoupling between hydrological time-series multi-source driving factors and response processes, outputting a decoupled potential state sequence, corresponding to the causal constraint encoder and decoupled potential state in the architecture diagram of Figure 2. Module.

[0093] S6: Time Series Predictive Inference

[0094] To conduct temporal prediction inference, a temporal prediction network is constructed, using a spatiotemporal graph convolutional network (STGCN). The decoupled latent state sequence obtained in step S5 is then used. The dynamic features of the graph nodes are mapped, and the adjacency matrix of the hydrological causal graph obtained in step S4 is used as the weight constraint of the graph convolution operator. Temporal dependencies are captured by the gated convolution units inside STGCN, and spatial causal responses of different hydrological variables are captured by the graph convolution layer. The measured water level and flow data of the river control section in the future period are used as labels, and the MAE of formula (8) is used as the core loss function for model optimization.

[0095]

[0096] Where N is the number of training samples; To predict the number of cross sections; This represents the predicted water level / flow rate at the k-th cross-section of the n-th sample. The corresponding measured value; the recursive multi-step prediction strategy uses the single-step prediction result of the time series prediction network as the input feature for the next prediction, and simultaneously integrates the historical time series data corresponding to the time delay characteristics determined in step S3, and iteratively executes to obtain the hydrological state sequence within the preset prediction period. The maximum prediction period is set to 1d~30d, and the single-step prediction step size is consistent with the time resolution of the original hydrological data.

[0097] Complete model training and optimization; adopt a recursive multi-step prediction strategy, use the single-step prediction results of the trained time series prediction network as the input features for the next prediction, and simultaneously integrate the historical time series data corresponding to the time delay characteristics determined in step S3, iteratively execute the prediction inference process, and obtain the hydrological state sequence of the plain river network within the preset forecast period, corresponding to the time series prediction network module in the architecture diagram of Figure 2.

[0098] S7: Output Results

[0099] Based on the prediction and reasoning results of step S6, the prediction results of the core hydrological indicators of future water level and flow in the plain river network within the forecast period are output. The core driving causal factors and contribution of the corresponding prediction results are output simultaneously. The prediction accuracy is verified by formulas (9), (8), and (10).

[0100]

[0101]

[0102] in, The result is the mean of all measured values; the output result synchronously includes the core driving causal factors and contribution of the corresponding prediction result, realizing the interpretability of the hydrological model reasoning process.

[0103] In summary, the entire inference process of the plain river network hydrological model is completed, corresponding to the output layer module in the architecture diagram of Figure 2.

[0104] The method of this invention achieves relative improvements of over 20% in three core hydrological prediction indicators: root mean square error (RMSE), mean absolute error (MAE), and Nash efficiency coefficient (NSE). Specifically, compared to the traditional hydrological model (Model D), the RMSE reduction rate is as high as 27.4%, and the MAE reduction rate is as high as 30.0%, fully demonstrating the superior prediction performance of this invention in the complex hydrological environment of plain river networks. Compared to models without causal constraints and models without physical constraints, this invention effectively removes confounding factors from multi-source hydrological data through causal decoupling, reducing redundant information interference.

[0105] This invention identifies the true driving relationships of hydrological elements by integrating temporal causal discovery with physical mechanism constraints, achieves decoupling representation of multi-source coupled hydrological data through causal constraint encoder, and combines temporal deep learning network and recursive multi-step prediction strategy to significantly improve the inference and prediction accuracy and interpretability of core hydrological indicators while satisfying the hydrological physical mechanism of plain river network.

[0106] Example 2: This example proposes a reasoning system for a plain river network hydrological model based on causal decoupling, including:

[0107] The data acquisition module is configured to perform the following process: acquire multi-source hydrological time series data of the plain river network;

[0108] The data preprocessing and periodic feature screening module is configured to perform the following processes: standardize the multi-source hydrological time series, divide it into training set, validation set and test set, use the sliding window method to screen potential periodic features, and determine the optimal historical window length.

[0109] The temporal causality discovery module is configured to perform the following process: based on the preprocessed hydrological time series and the optimal historical window length, it conducts temporal causality identification, mines the statistical causal relationships and corresponding time lag characteristics between hydrological elements, incorporates the constraints of the hydrological physical mechanism of the plain river network, filters out false associations that do not conform to hydrological laws, and obtains effective causal relationships with physical rationality.

[0110] The hydrological cause-effect graph construction and prior output module is configured to perform the following process: based on the results of temporal causal relationship identification, construct a hydrological cause-effect graph of the plain river network that includes the influence relationship and time lag characteristics, and output the causal relationship and time lag prior;

[0111] The causal decoupling representation learning module is configured to perform the following process: construct a causal constraint encoder based on the hydrological causal graph, embed causal relationships and time-delay priors into the structure mask and parent node constraints of the encoder, and map the preprocessed hydrological data into decoupled latent states through the encoder to achieve causal decoupling in the hidden space;

[0112] The temporal prediction inference module is configured to perform the following process: conduct temporal prediction inference, input the decoupled potential state sequence into the spatiotemporal graph convolutional network, capture the temporal dependencies and periodic characteristics of the hydrological process, and infer the future hydrological state through a recursive multi-step prediction strategy.

[0113] The results output module is configured to perform the following process: output the predicted results of future water levels and flow rates in the plain river network.

[0114] Example 3: This example proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the control method as described in this invention.

[0115] Example 4: This example proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed, it implements the steps of the method described in this invention.

[0116] It should be noted that the processing flow of embodiments 2-4 corresponds to the specific steps of the method provided in embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in embodiment 1 of the present invention.

[0117] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0118] The specific implementation schemes described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific implementation schemes of the present invention and are not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.

Claims

1. A reasoning method for a plain river network hydrological model based on causal decoupling, characterized in that, Includes the following steps: S1. Collect multi-source hydrological time series data of plain river network; S2. Standardize and preprocess the multi-source hydrological time series, divide it into training set, validation set and test set, use the sliding window method to initially screen potential periodic features, and determine the optimal historical window length. S3. Based on the preprocessed hydrological time series and the optimal historical window length, conduct temporal causal relationship identification, mine the statistical causal relationship and corresponding time lag characteristics between hydrological elements, incorporate the constraints of the hydrological physical mechanism of the plain river network, filter out false associations that do not conform to hydrological laws, and obtain effective causal relationships with physical rationality. S4. Based on the results of temporal causal relationship identification, construct a hydrological causal diagram of the plain river network that includes the influence relationship and time lag characteristics, and output the causal relationship and time lag prior. S5. Construct a causal constraint encoder based on the hydrological causal graph, embed causal relationships and time-delay priors into the structure mask and parent node constraints of the encoder, and map the preprocessed hydrological data into decoupled potential states through the encoder to achieve causal decoupling in the hidden space. S6. Conduct temporal prediction reasoning, input the decoupled potential state sequence into the spatiotemporal graph convolutional network, capture the temporal dependency and periodic characteristics of the hydrological process, and infer the future hydrological state through a recursive multi-step prediction strategy. S7. Output the predicted results of future water level and flow in the plain river network.

2. The method according to claim 1, characterized in that, The standardization preprocessing described in step S2 employs two methods: Min-Max normalization and Z-score normalization. Min-Max normalization linearly maps the original hydrological data to the [0,1] interval, eliminating the influence of dimensions. Specifically, it is calculated using the following formula: ， in, For the first The original values ​​of the hydrological variables at time t; The value is the normalized value; , The first The maximum and minimum values ​​of each hydrological variable throughout the entire time series; Z-score standardization is used to transform the raw hydrological data into a standard normal distribution with a mean of 0 and a standard deviation of 1, specifically calculated using the following formula: ， in, These are the standardized values; For the first The mean of the time series of several hydrological variables; For the first The standard deviation of the time series of hydrological variables; the training set, validation set and test set are divided in a 7:1.5:1.5 ratio according to time order; the sliding window method sets the window scale to cover 1d, 7d, 15d, 30d, 90d and 180d, and identifies potential periodic characteristics by calculating the autocorrelation coefficient and partial correlation coefficient of the sequence, and determines the optimal historical window length for time series modeling.

3. The method according to claim 1, characterized in that, In step S3, the partial correlation coefficient is used to measure the conditional independence of two hydrological variables under a given set of conditions. The partial correlation coefficient is calculated using the following formula: ， in, Let X be the partial correlation coefficient between variables X and Y under condition set Z; To remove variables from condition set Z The set after; when When X and Y are conditionally independent, we accept the null hypothesis, where Φ−1 is the inverse function of the standard normal distribution and α is the significance level, which is 0.

05. The MCI criterion is used to calculate the strength of causal effects, specifically using the following formula: ， in, Let variable X be a time delay of Y The estimated causal effect values; for The set of parent nodes contains the lagged terms of Y itself and the lagged terms of other potential driving factors; when When a causal relationship passes the significance test and meets the physical mechanism constraints, it is determined to be a valid causal relationship.

4. The method according to claim 1, characterized in that, The hydrological causal graph of the plain river network described in step S4 uses hydrological elements such as precipitation, upstream water inflow, gate gravity drainage, pumping station drainage, real-time tides, river water level, and flow as nodes, the effective causal relationships obtained in step S3 as directed edges, and the corresponding time delay characteristics as edge weights to construct a directed acyclic graph structure; the causal relationships and time delay priors include the causal relationship binary adjacency matrix, the parent node dependency relationship of each node, and the time delay parameters.

5. The method according to claim 1, characterized in that, In step S5, a structural mask is constructed based on the binary adjacency matrix of the plain river network hydrological causal graph to constrain the neuron connections of the encoder, specifically calculated using the following formula: ， in, The structure mask weight matrix of the encoder; Given a binary adjacency matrix of a hydrological cause-effect graph, if the nodes... It is a node The parent node, then Otherwise, it is 0; This is the original weight matrix of the encoder; For the Hadamardi, ensure that information is transmitted only between nodes with causal relationships; The feature propagation process of each layer of the causal constraint encoder is calculated using the following formula: ， in, Let L be the network weights of the encoder's layer l. It is a non-linear activation function. This is the feature output of the encoder's l-th layer; Based on a structural causal model, the preprocessed hydrological data is compared with historical time series windows. Mapping to decoupled potential states The specific calculation is as follows: ， in, for The decoupled latent state vector at time step; This is the weighted adjacency matrix of the hydrological causal graph, corresponding to the intensity of causal effects; It is the identity matrix; , It is an element-wise nonlinear invertible transformation function; For exogenous noise vectors that follow a standard normal distribution, The hidden layer neurons of the causal constraint encoder strictly follow the directed acyclic structure of the hydrological causal graph, allowing only information transmission from parent nodes to child nodes and prohibiting reverse information flow.

6. The method according to claim 1, characterized in that, The spatiotemporal graph convolutional network described in step S6 receives the decoupled latent state output in step S5. As node features, and using the hydrological causal graph constructed in step S4 as spatial topological constraints, the spatiotemporal coupling characteristics between hydrological elements are captured; the model training uses mean absolute error as the core loss function, as shown in the following formula: ， Where N is the number of training samples; To predict the number of cross sections; This represents the predicted water level / flow rate at the k-th cross-section of the n-th sample. The corresponding measured value; the recursive multi-step prediction strategy uses the single-step prediction result of the time series prediction network as the input feature for the next prediction, and simultaneously integrates the historical time series data corresponding to the time delay characteristics determined in step S3, and iteratively executes to obtain the hydrological state sequence within the preset prediction period. The maximum prediction period is set to 1d~30d, and the single-step prediction step size is consistent with the time resolution of the original hydrological data.

7. The method according to claim 1, characterized in that, The accuracy evaluation indicators of the prediction results in step S7 include root mean square error, mean absolute error, and coefficient of determination; the root mean square error is calculated using the following formula: ， The coefficient of determination is calculated using the following formula: ， in, The result is the mean of all measured values; the output result synchronously includes the core driving causal factors and contribution of the corresponding prediction result, realizing the interpretability of the hydrological model reasoning process.

8. A reasoning system for a plain river network hydrological model based on causal decoupling, characterized in that, include: The data acquisition module is configured to perform the following process: acquire multi-source hydrological time series data of the plain river network; The data preprocessing and periodic feature screening module is configured to perform the following processes: standardize the multi-source hydrological time series, divide it into training set, validation set and test set, use the sliding window method to screen potential periodic features, and determine the optimal historical window length. The temporal causality discovery module is configured to perform the following process: based on the preprocessed hydrological time series and the optimal historical window length, it conducts temporal causality identification, mines the statistical causal relationships and corresponding time lag characteristics between hydrological elements, incorporates the constraints of the hydrological physical mechanism of the plain river network, filters out false associations that do not conform to hydrological laws, and obtains effective causal relationships with physical rationality. The hydrological cause-effect graph construction and prior output module is configured to perform the following process: based on the results of temporal causal relationship identification, construct a hydrological cause-effect graph of the plain river network that includes the influence relationship and time lag characteristics, and output the causal relationship and time lag prior; The causal decoupling representation learning module is configured to perform the following process: construct a causal constraint encoder based on the hydrological causal graph, embed causal relationships and time-delay priors into the structure mask and parent node constraints of the encoder, and map the preprocessed hydrological data into decoupled latent states through the encoder to achieve causal decoupling in the hidden space; The temporal prediction inference module is configured to perform the following process: conduct temporal prediction inference, input the decoupled potential state sequence into the spatiotemporal graph convolutional network, capture the temporal dependencies and periodic characteristics of the hydrological process, and infer the future hydrological state through a recursive multi-step prediction strategy. The results output module is configured to perform the following process: output the predicted results of future water levels and flow rates in the plain river network.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 7.

Images