Intelligent control method and system based on AI hydrological telemetry

Abstract

The application relates to the technical field of hydrological telemetry, in particular to an intelligent control method and system based on AI hydrological telemetry, which constructs a causal diagram model by acquiring multi-source hydrological telemetry data, identifies a causal dependence relationship to obtain a causal structure representation, performs reverse causal path tracking and root cause contribution analysis on the basis of the representation, generates an intervention control scheme accordingly, and updates the causal diagram model and performs parameter correction by using feedback data after executing the scheme, so that the accuracy and self-adaptability of hydrological system anomaly intervention are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of hydrological telemetry technology, and in particular to an intelligent control method and system based on AI hydrological telemetry. Background Technology

[0002] Currently, the field of hydrological telemetry and control generally adopts monitoring methods based on threshold judgment and univariate statistical analysis. The conventional approach involves deploying sensors for water level, flow rate, and rainfall at key cross-sections, setting fixed thresholds to trigger alarms, and then relying on human experience or simple correlation analysis to trace the source of anomalies. Finally, control commands are generated based on a pre-defined rule base or physical model. This model relies on a large amount of historical data to establish statistical regression equations or uses distributed hydrological models based on physical processes for simulation and extrapolation. Some systems introduce machine learning algorithms to classify and identify anomalies, but these are mostly limited to surface feature mapping and do not delve into the intrinsic mechanisms between variables.

[0003] The aforementioned mainstream methods suffer from two significant drawbacks. Firstly, threshold rules and statistical models essentially only characterize the correlation between variables, not causal relationships. Secondly, when hydrological systems exhibit nonlinear coupling, time-delay propagation, and multivariate interactions, correlation coefficients can be misled by spurious associations or confounding factors, leading to frequent abnormal alarms and high false alarm rates. For instance, there are unobserved intermediate variables between upstream rainfall and downstream water level fluctuations; simple statistical regression struggles to distinguish between direct and indirect effects, thus failing to accurately pinpoint the true root cause. This results in interventions that are often delayed or deviate from the actual problem. Thirdly, while physical models are based on hydrodynamic principles, parameter calibration is complex and heavily reliant on system steady-state assumptions. In extreme events or dynamic scenarios such as engineering scheduling, model parameters quickly become inaccurate due to their inability to adaptively adjust. More critically, existing methods generally lack closed-loop feedback mechanisms: after control commands are executed, changes in system state are not effectively utilized to correct model structure and parameters, leading to a continuous accumulation of deviations between the model and the physical system, and a sharp decline in control accuracy over time. This open-loop or semi-open-loop control mode struggles to cope with the time-varying and uncertainties of hydrological systems, limiting the reliability of intelligent decision-making. Summary of the Invention

[0004] The embodiments of the present invention provide an intelligent control method and system based on AI hydrological telemetry, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an intelligent control method based on AI hydrological telemetry, comprising: Acquire multi-source hydrological telemetry data, construct a causal graph model, and use a constrained optimization algorithm to identify causal dependencies, time delay propagation characteristics, and nonlinear coupling modes among variables in the multi-source hydrological telemetry data, thereby obtaining a causal structure representation that includes nodes, directed edges, and causal strength weights; Based on the aforementioned causal structure representation, reverse causal path tracing and multi-level propagation analysis are performed on the detected abnormal observation variables to identify all upstream root cause nodes. The causal contribution of each root cause node to the anomaly is quantified by intervention effect estimation and counterfactual comparison, forming a root cause node set and its corresponding contribution distribution. Based on the root cause node set and the causal contribution distribution, combined with hydrological control constraints and system physical laws, an intervention and control scheme is generated. The intervention and control scheme is executed and feedback observation data is collected after execution. The feedback observation data is used to update the structural topology and causal strength weights of the causal graph model online, and the parameters are corrected for the deviation between the actual intervention effect and the predicted effect.

[0006] Multi-source hydrological telemetry data are acquired, a causal graph model is constructed, and a constrained optimization algorithm is used to identify causal dependencies, time-delay propagation characteristics, and nonlinear coupling modes among variables from the multi-source hydrological telemetry data, resulting in a causal structure representation including: Mutual information calculation and conditional independence test are performed on multi-source hydrological telemetry data to quantify the statistical correlation strength and conditional dependence between any two variables. Based on the statistical correlation strength and conditional dependence, gradient iteration is performed by applying acyclic constraints and sparsity penalties to obtain an initial directed acyclic graph. For each pair of variables corresponding to a directed edge in the initial directed acyclic graph, cross-correlation analysis is used to search for the number of lag steps that maximize the causal transmission effect within multiple candidate time lag intervals. The number of lag steps and the strength of the causal transmission effect are then labeled on the corresponding directed edges to obtain a time lag labeling graph containing time lag information. For each directed edge in the time-delay annotation graph, the historical observations of the source node at the corresponding lag step and the current observations of the target node are extracted to form a sample pair. The mapping pattern from the historical observations of the source node to the current observations of the target node is fitted by a nonlinear regression method to characterize the nonlinear coupling mode between variables. The topological structure of the time-delay annotation map, the number of lag steps, and the fitted nonlinear mapping pattern are integrated into a causal structure representation, which simultaneously encodes causal dependencies, time-delay propagation characteristics, and nonlinear coupling modes.

[0007] Based on the statistical correlation strength and the conditional dependency, gradient iteration is performed by applying acyclic constraints and sparsity penalties to obtain the initial directed acyclic graph, including: An adjacency matrix is ​​constructed to represent the directed connections between variables, and the elements of the adjacency matrix are initialized and assigned values ​​according to the statistical association strength. The trace of the matrix exponential function is calculated based on the adjacency matrix to form a loop quantization index. The cumulative amount of the absolute value of the elements of the adjacency matrix and the loop quantization index are respectively set as the acyclicity constraint term and the sparsity penalty term. A joint loss function is constructed by combining the fitting error of the conditional dependence degree. The gradient of the joint loss function is calculated for the elements of the adjacency matrix. The element values ​​of the adjacency matrix are iteratively corrected according to the opposite direction of the gradient. During the correction process, the weight coefficient of the sparsity penalty term is adaptively determined according to the descent rate of the loop quantization index. In the iterative stage where the loop quantization index decreases slowly, the weight of the sparsity penalty term is increased to accelerate the edge pruning. Repeatedly calculate the loop quantization index, the gradient, and the adjacency matrix correction until the loop quantization index converges to the zero neighborhood and the joint loss function no longer decreases. Then, map the final adjacency matrix to generate the initial directed acyclic graph.

[0008] Based on the aforementioned causal structure representation, reverse causal path tracing and multi-level propagation analysis are performed on the detected anomalous observed variables to identify all upstream root cause nodes. The causal contribution of each root cause node to the anomaly is quantified through intervention effect estimation and counterfactual comparison, forming a root cause node set and its corresponding contribution distribution, including: Based on the endpoint matching rules of directed edges, all incoming edges with the endpoint of the abnormal observation variable are filtered from the causal structure representation, the association is established, and the starting node of the incoming edge and the lag step number marked on the edge are recorded. For the starting node of the incoming edge, continue to parse its incoming edge, backtrack layer by layer along the direction of the incoming edge and mark the level depth until the node no longer has an incoming edge, and determine the node without an incoming edge as the upstream root cause node based on the feature of missing incoming edges. A directed path from each upstream root cause node to the abnormal observation variable is generated by depth-first traversal. The lag steps of each edge on the directed path are superimposed and transferred along the path direction to obtain the end-to-end time delay. The nonlinear mapping rules of each edge on the directed path are combined in the connection order to obtain the path cascade transfer function. For each upstream root cause node, the observed value of the node is reconstructed into the historical normal median value based on the historical distribution characteristics to construct the intervention state. The propagation time offset of the intervention state is inferred based on the end-to-end time delay corresponding to the node. The response value of the intervention state propagating through each directed path is calculated using the path concatenation transfer function, and the response values ​​are fused according to the propagation intensity of each directed path to obtain the counterfactual estimate of the abnormal observation variable. Calculate the deviation between the actual observed value and the counterfactual estimated value of the abnormal observed variable, and map the deviation to the causal contribution of the corresponding upstream root cause node through a difference metric; Sort and organize all upstream root cause nodes and their corresponding causal contributions according to the causal contribution degree, forming a root cause node set and its corresponding contribution degree distribution.

[0009] For each upstream root cause node, the observed values ​​of the upstream root cause node are reconstructed into historical normal median values ​​based on historical distribution characteristics to construct the intervention state. The propagation time offset of the intervention state is inferred based on the end-to-end time delay corresponding to the upstream root cause node, including: For each upstream root cause node, the observation sequence of the upstream root cause node in the normal operation interval is extracted from the historical time window, and the historical normal median value is obtained by performing quantile statistics on the observation sequence in chronological order. Replace the actual observed value of the upstream root cause node at the time of the anomaly with the historical normal median value, and construct the intervention state observation value of the upstream root cause node as a counterfactual benchmark. Extract the end-to-end time delay on the directed path from the upstream root cause node to the abnormal observation variable, and use the end-to-end time delay as the propagation delay step number of the intervention state observation value from the upstream root cause node to the abnormal observation variable. Based on the propagation delay steps, the effective time of the intervention state observation is shifted backward by the time length corresponding to the propagation delay steps to obtain the arrival time of the intervention state at the abnormal observation variable. The time difference between the arrival time and the time of the abnormal occurrence of the abnormal observation variable is determined as the propagation time offset of the intervention state.

[0010] Based on the root cause node set and the causal contribution distribution, combined with hydrological control constraints and system physical laws, the following intervention and control schemes are generated: Identify and index each upstream root cause node from the root cause node set and associate it with its corresponding causal contribution value in the causal contribution distribution, and construct an intervention priority ranking of the root cause nodes according to the causal contribution value. For each upstream root cause node in the intervention priority ranking, the hydrological control variables associated with the upstream root cause node and the physical adjustable range of the hydrological control variables are identified and established from the causal structure representation. The physical adjustable range is marked and constructed as the initial control space of the upstream root cause node. Obtain the hydrological control constraints of the hydrological system, and decompose and reorganize the hydrological control constraints into the constraint boundaries of the initial control space of each upstream root node; Based on the constraint boundary, the initial control space of each upstream root cause node is iteratively shrunk and adaptively adjusted to generate a feasible control space for each upstream root cause node. Within the feasible control space of the upstream root cause node, the control adjustment value is optimized and locked. The path cascade transfer function from the upstream root cause node to the abnormal observation variable and the end-to-end time delay are used to calculate the expected correction effect of the control adjustment value on the abnormal observation variable through the causal path. The control adjustment values ​​and expected correction effects of each upstream root cause node are sequentially arranged and aggregated into an intervention control scheme according to the intervention priority.

[0011] The feedback observation data is used to update the structural topology and causal strength weights of the causal graph model online, and parameter correction is performed to address the deviation between the actual intervention effect and the predicted effect, including: The time-series observation values ​​of each monitoring node after the intervention are analyzed from the feedback observation data. The time-series observation values ​​are then aligned with the historical observation values ​​of the corresponding nodes before the intervention, and the change in the observation values ​​of each node before and after the intervention is calculated. The observed changes are compared node by node with the expected correction effect preset in the intervention and control plan, and the deviation between the actual intervention effect and the expected correction effect is calculated. Based on the deviation value, identify the source of prediction error of the propagation path in the causal graph model. For causal edges where the prediction error exceeds the allowable range, extract the node pairs connected by the causal edge and the causal strength weight labeled on the causal edge. Based on the actual response relationship of node pairs in the feedback observation data, the causal strength weight is numerically corrected, and the corrected causal strength weight is written back to the corresponding causal edge of the causal graph model. Detect whether there are any new causal relationships in the feedback observation data that are not covered by the causal graph model. If there are new causal relationships, insert new causal edges into the structural topology of the causal graph model and label the initial causal strength weights. The updated causal strength weights and structural topology are synchronized to the causal graph model to complete the online update and parameter correction of the causal graph model.

[0012] A second aspect of this invention provides an intelligent control system based on AI hydrological telemetry, comprising: The causal construction unit is used to acquire multi-source hydrological telemetry data, construct a causal graph model, and identify causal dependencies, time delay propagation characteristics and nonlinear coupling modes among variables from the multi-source hydrological telemetry data through a constrained optimization algorithm, thereby obtaining a causal structure representation containing nodes, directed edges and causal strength weights. The root cause analysis unit is used to perform reverse causal path tracing and multi-level propagation analysis on the detected abnormal observation variables based on the causal structure representation, identify all upstream root cause nodes, and quantify the causal contribution of each root cause node to the anomaly through intervention effect estimation and counterfactual comparison, forming a root cause node set and corresponding contribution distribution. An intervention generation unit is used to generate an intervention control scheme based on the root cause node set and the causal contribution distribution, combined with hydrological control constraints and system physical laws. The feedback update unit is used to execute the intervention control scheme and collect feedback observation data after execution. It uses the feedback observation data to update the structural topology and causal strength weights of the causal graph model online, and at the same time corrects the deviation between the actual intervention effect and the predicted effect.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] Accurate identification of causal dependencies, time-delay propagation characteristics, and nonlinear coupling modes among variables in hydrological systems significantly improves the accuracy of anomaly diagnosis. By constructing a causal graph model and applying constrained optimization algorithms, a clear causal structure representation can be extracted from massive multi-source telemetry data, avoiding spurious associations caused by traditional correlation analysis. When anomaly observed variables are detected, reverse causal path tracing and multi-level propagation analysis are used to accurately locate all upstream root cause nodes. Furthermore, the causal contribution of each root cause is quantified through intervention effect estimation and counterfactual comparison, forming a root cause set with clear priorities and contribution distribution, significantly shortening the anomaly tracing time.

[0016] Based on the root cause node and contribution distribution, the intervention and control scheme generated by combining hydrological control constraints and system physical laws is highly targeted and feasible. The intervention scheme directly acts on the actual pathogenic variables, avoiding blind adjustment of irrelevant parameters, thereby reducing control costs and minimizing interference with the normal operation of the system. Simultaneously, the intervention effect is quantitatively evaluated using causal strength weights to ensure the scheme's optimality within the physically feasible range, effectively improving the stability and response efficiency of the hydrological system.

[0017] The causal graph model is updated online using feedback observation data after implementation, including dynamic correction of structural topology and causal strength weights, enabling the model to continuously adapt to changes in the hydrological environment. Deviations between actual intervention effects and predicted effects are automatically compensated through parameter correction, forming a closed-loop self-optimization mechanism and reducing the need for manual intervention. This method can quickly restore system equilibrium after anomalies occur and accumulate causal knowledge from historical data, gradually improving the predictive accuracy and robustness of future intervention strategies. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the intelligent control method based on AI hydrological telemetry according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the online update and parameter correction process of the causal graph model in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the intelligent control method based on AI hydrological telemetry according to an embodiment of the present invention.

[0022] Intelligent control methods based on AI hydrological remote sensing include: Acquire multi-source hydrological telemetry data, construct a causal graph model, and use a constrained optimization algorithm to identify causal dependencies, time delay propagation characteristics, and nonlinear coupling modes among variables in the multi-source hydrological telemetry data, thereby obtaining a causal structure representation that includes nodes, directed edges, and causal strength weights; Based on the aforementioned causal structure representation, reverse causal path tracing and multi-level propagation analysis are performed on the detected abnormal observation variables to identify all upstream root cause nodes. The causal contribution of each root cause node to the anomaly is quantified by intervention effect estimation and counterfactual comparison, forming a root cause node set and its corresponding contribution distribution. Based on the root cause node set and the causal contribution distribution, combined with hydrological control constraints and system physical laws, an intervention and control scheme is generated. The intervention and control scheme is executed and feedback observation data is collected after execution. The feedback observation data is used to update the structural topology and causal strength weights of the causal graph model online, and the parameters are corrected for the deviation between the actual intervention effect and the predicted effect.

[0023] In one optional implementation, multi-source hydrological telemetry data is acquired, a causal graph model is constructed, and a constrained optimization algorithm is used to identify causal dependencies, time-delay propagation characteristics, and nonlinear coupling modes among variables from the multi-source hydrological telemetry data, resulting in a causal structure representation including: Mutual information calculation and conditional independence test are performed on multi-source hydrological telemetry data to quantify the statistical correlation strength and conditional dependence between any two variables. Based on the statistical correlation strength and conditional dependence, gradient iteration is performed by applying acyclic constraints and sparsity penalties to obtain an initial directed acyclic graph. For each pair of variables corresponding to a directed edge in the initial directed acyclic graph, cross-correlation analysis is used to search for the number of lag steps that maximize the causal transmission effect within multiple candidate time lag intervals. The number of lag steps and the strength of the causal transmission effect are then labeled on the corresponding directed edges to obtain a time lag labeling graph containing time lag information. For each directed edge in the time-delay annotation graph, the historical observations of the source node at the corresponding lag step and the current observations of the target node are extracted to form a sample pair. The mapping pattern from the historical observations of the source node to the current observations of the target node is fitted by a nonlinear regression method to characterize the nonlinear coupling mode between variables. The topological structure of the time-delay annotation map, the number of lag steps, and the fitted nonlinear mapping pattern are integrated into a causal structure representation, which simultaneously encodes causal dependencies, time-delay propagation characteristics, and nonlinear coupling modes.

[0024] After acquiring multi-source hydrological telemetry data, the raw data is first preprocessed, including missing value imputation, outlier removal, and time series alignment, to ensure the comparability of observed variables such as precipitation, runoff, water level, soil moisture content, and evapotranspiration from different sensor stations under a unified time reference. After preprocessing, mutual information is calculated between all variables to quantify the strength of nonlinear statistical associations between variables. Let variable X... i With X j The mutual information is I(X) i X j This reflects the amount of information shared between the two variables, is not constrained by the linear assumption, and can capture nonlinear correlations in complex hydrological processes. Based on mutual information calculation, a conditional independence test is further performed, that is, given a third variable set Z, testing X... i With X jWhether conditional independence exists is crucial for distinguishing between genuine causal relationships and spurious relationships caused by common driving variables. The conditional independence test employs a kernel-based nonparametric strategy to accommodate the diversity and non-Gaussian nature of hydrological data distribution.

[0025] Based on the mutual information matrix and conditional independence test results, the learning of causal graph structure is transformed into a constrained optimization problem. A continuously differentiable feature representation of the directed acyclic graph is introduced, embedding the acyclic constraint into the objective function in the form of equality constraints. Simultaneously, a sparsity penalty term is applied to avoid overfitting and improve the interpretability of the graph structure. Let W be the weighted adjacency matrix of the causal graph. The objective function comprehensively considers three components: data fitting loss, acyclic penalty, and sparsity regularization. The optimal adjacency matrix is ​​obtained through gradient iteration. .right Binarization is performed based on a threshold, and statistically significant directed edges are retained to obtain an initial directed acyclic graph (DAG). Each node in this DAG corresponds to a hydrological telemetry variable, and each directed edge represents a direct causal dependency between a pair of variables. The edge weight reflects the relative strength of the causal dependency.

[0026] For each variable pair corresponding to a directed edge in the initial directed acyclic graph, a time delay identification process is performed. In hydrological systems, the transmission of causal signals generally involves time delays; for example, the impact of upstream precipitation events on downstream runoff often lags by hours or even days. For directed edge X... i X j Within the preset candidate time lag interval Inside, calculate X one by one. i Different lag steps The historical observation sequence and X j The cross-correlation coefficient between the current observation sequences. The calculation of the cross-correlation coefficient covers both positive and negative lag directions to identify the optimal number of lag steps. Even the lag value for maximizing the causal transmission effect. The maximum cross-correlation coefficient, representing the strength of the causal transmission effect, is labeled on the corresponding directed edge of the directed acyclic graph, forming a time-delay labeling graph. The time-delay labeling graph maintains the same topological structure as the initial directed acyclic graph, but each directed edge additionally carries two attributes: the number of lag steps and the transmission strength. This allows the causal graphical model to accurately describe the temporal dynamics of signal propagation during hydrological processes.

[0027] Based on the time-delay annotation diagram, the nonlinear coupling mode is characterized for each directed edge. For directed edge X... i X j Extract X i Lag steps Using the historical observation sequence as input features, with Xj The current observation sequence is used as the regression target to construct a set of sample pairs. A nonlinear regression method is then used to fit these sample pairs; specifically, Gaussian process regression, kernel ridge regression, or lightweight neural networks can be selected flexibly based on the sample size and computational resources. Gaussian process regression can simultaneously provide the predicted mean and uncertainty estimate, making it suitable for scenarios with small datasets and requiring quantified prediction confidence; kernel ridge regression has high computational efficiency for medium-sized datasets; and lightweight neural networks are suitable for capturing more complex high-dimensional nonlinear mapping relationships. The fitted nonlinear mapping function f... ij Describe X i How does the historical state nonlinearly affect X? j The current state characterizes the dynamic coupling between variables under time-delay conditions. The above nonlinear regression fitting process is repeated for all directed edges in the time-delay labeled graph to establish a corresponding nonlinear mapping model for each edge.

[0028] After completing the time delay identification and nonlinear coupling pattern fitting, the topological structure of the time delay annotation graph and the number of lag steps for each directed edge are determined. The strength of causal transmission effect and the fitted nonlinear mapping function f ij The data is integrated to form a complete causal structure representation. This causal structure representation encodes three types of information simultaneously using a unified data structure: first, causal dependencies, expressed by the topological structure and edge weights of a directed acyclic graph, clearly identifying which variable pairs have direct causal influences and their relative strengths; second, time-delay propagation characteristics, expressed by the lag steps and propagation strength labeled on each directed edge, reflecting the propagation time characteristics of hydrological signals in the spatial and physical process links; and third, nonlinear coupling patterns, expressed by the nonlinear mapping function corresponding to each directed edge, capturing the complex dynamic relationships between variables that transcend the linear assumption. The organic integration of these three types of information enables the causal graph model to have a comprehensive ability to represent the causal mechanisms of hydrological systems, providing a reliable structured knowledge foundation for subsequent anomaly root cause tracing and intervention control scheme generation.

[0029] In practical engineering applications, multi-source hydrological telemetry data often originates from sensors of different accuracy levels, and the sampling frequencies vary. To ensure the robustness of the causal graph construction, adaptive bandwidth estimation is introduced in the mutual information calculation stage to address the differences in the distribution characteristics of different variables. In the conditional independence testing stage, a multiple test correction strategy is employed to control the false positive rate and avoid introducing false causal edges due to an excessive number of statistical tests. During the gradient iterative solution process, a strategy of gradually tightening the acyclicity penalty is adopted. First, relatively relaxed constraints are used to quickly converge to a reasonable neighborhood, and then the constraints are gradually strengthened to obtain a final solution that strictly satisfies acyclicity, thus balancing solution efficiency and result quality. The entire causal graph construction process supports incremental updates. When new telemetry data is continuously integrated, the local causal structure can be corrected without recalculating the entire dataset, maintaining the model's adaptability to the dynamic evolution of the hydrological system.

[0030] In one optional implementation, the initial directed acyclic graph is obtained by iteratively solving for the statistical correlation strength and the conditional dependency by applying acyclicity constraints and sparsity penalties, including: An adjacency matrix is ​​constructed to represent the directed connections between variables, and the elements of the adjacency matrix are initialized and assigned values ​​according to the statistical association strength. The trace of the matrix exponential function is calculated based on the adjacency matrix to form a loop quantization index. The cumulative amount of the absolute value of the elements of the adjacency matrix and the loop quantization index are respectively set as the acyclicity constraint term and the sparsity penalty term. A joint loss function is constructed by combining the fitting error of the conditional dependence degree. The gradient of the joint loss function is calculated for the elements of the adjacency matrix. The element values ​​of the adjacency matrix are iteratively corrected according to the opposite direction of the gradient. During the correction process, the weight coefficient of the sparsity penalty term is adaptively determined according to the descent rate of the loop quantization index. In the iterative stage where the loop quantization index decreases slowly, the weight of the sparsity penalty term is increased to accelerate the edge pruning. Repeatedly calculate the loop quantization index, the gradient, and the adjacency matrix correction until the loop quantization index converges to the zero neighborhood and the joint loss function no longer decreases. Then, map the final adjacency matrix to generate the initial directed acyclic graph.

[0031] After obtaining the statistical correlation strength and conditional dependence between each pair of variables, this statistical information needs to be transformed into a directed acyclic graph structure with clear directionality and guaranteed acyclicity. To this end, a graph structure learning framework based on continuous optimization is introduced, using the adjacency matrix as the optimization variable, and solving for the optimal graph structure through gradient iteration while satisfying the constraints of the directed acyclic graph.

[0032] Construct an n×n real-valued adjacency matrix A, where n is the total number of hydrological telemetry variables, and the matrix elements A pqThis represents the directed connection strength from the p-th variable to the q-th variable. During initialization, the elements of the adjacency matrix are assigned values ​​based on pre-calculated statistical association strengths: if the association strength between two variables exceeds a preset threshold, the corresponding element is initialized to the normalized value of that association strength; if the association strength is below the threshold, the corresponding element is initialized to a small value close to zero. This initialization strategy allows the optimization process to start from a point consistent with the statistical characteristics of the data, helping to accelerate subsequent gradient iteration convergence while reducing the risk of getting trapped in local optima. Diagonal elements are forcibly set to zero to exclude autoregressive connections between variables and themselves.

[0033] When constructing the joint loss function, three constraints and objectives need to be considered simultaneously. The path quantization index is constructed using the trace form of a matrix exponential function. Specifically, it is the element-wise squared matrix of the adjacency matrix. Calculate the matrix exponential function and take its trace to obtain the loop quantization index. The index is zero if and only if there are no directed cycles in the graph; when cycles exist, the index is strictly greater than zero, and the more cycles there are and the greater their weights, the higher the index value. This continuously differentiable cycle quantization form allows the acyclicity constraint to be incorporated into the gradient optimization framework, thus avoiding the high computational cost of discrete combinatorial search.

[0034] The sparsity penalty term is defined as the cumulative sum of the absolute values ​​of all elements in the adjacency matrix, i.e. Its function is to suppress the generation of redundant edges while preserving true causal connections, resulting in a sparse causal graph structure. The fitting error term is constructed based on the degree of conditional dependence, measuring how well the causal structure represented by the current adjacency matrix fits the conditional dependencies of variables in the observed data. Specifically, it is the sum of squared prediction residuals of each variable under the conditions of its parent node. Combining the three parts, the joint loss function is expressed as: ; in For the fitting error term, The penalty coefficient for the acyclicity constraint term. is the weighting coefficient of the sparsity penalty term. In the optimization process, the augmented Lagrange method is usually used to gradually increase the size to ensure that the final solution satisfies the strict acyclicity condition.

[0035] The gradient calculation is performed on each element of the adjacency matrix. The partial derivatives of the fitting error term, the loop quantization index, and the sparsity penalty term are calculated and summed to obtain the complete gradient matrix of the joint loss function with respect to A. In each iteration, the adjacent matrix elements are corrected along the opposite direction of the gradient: ,in For the learning rate, an adaptive learning rate strategy (such as the Adam optimizer) can be used to improve convergence stability. Near the zeros where the absolute value function is nondifferentiable, the gradient of the sparsity penalty term is approximated using subgradients to ensure the stability of numerical computation.

[0036] During the iterative correction process The value of h(A) is not fixed, but is adaptively adjusted according to the descent rate of the loop quantization index h(A). Specifically, after each optimization round, the difference between h(A) in the current round and the previous round is calculated. If the absolute value of this difference is lower than the preset descent rate threshold... If the current loop quantification index decreases slowly, then... According to a fixed multiplier Magnification enhances the pruning of redundant edges, prompting the graph structure to evolve towards sparser structures, indirectly accelerating loop elimination. When the rate of decrease in the loop quantification index returns to normal, The scaling operation is paused, and the current value is maintained to continue iteration. This adaptive mechanism can effectively deal with the "plateau" phenomenon in the optimization process and avoid the algorithm from stagnating for a long time on local structures with insufficient sparsity.

[0037] The convergence criterion includes two dimensions: first, the loop quantization index h(A) converges to the zero neighborhood, i.e. ,in The minimum positive number (e.g., 10) is a preset value. -8 This condition guarantees that the final graph structure satisfies the topological requirements of a directed acyclic graph; secondly, the joint loss function. No further significant decrease occurs in several consecutive iterations, meaning the absolute value of the loss difference between two adjacent iterations is below the convergence threshold. This condition ensures that the optimization has reached a stable state. Both conditions must be met simultaneously for the iteration to terminate.

[0038] After iterative convergence, the final adjacency matrix is... Thresholding is performed, removing edges whose absolute values ​​are below the edge pruning threshold. The elements are set to zero to eliminate small non-zero terms remaining from the numerical optimization process, resulting in a sparsified adjacency matrix. This sparsified adjacency matrix is ​​then mapped to the initial directed acyclic graph: non-zero elements... The directed edge from the p-th node to the q-th node in the graph corresponds to the edge weight, which is the absolute value of the element. The direction is determined by the sign of the element and the causal direction judgment rule. Thus, an initial directed acyclic graph with a clear topological structure and edge weights is obtained, providing a structural foundation for subsequent time delay propagation characteristic analysis and nonlinear coupled pattern recognition.

[0039] In real-world hydrological telemetry scenarios, the number of variables is large (e.g., encompassing multiple types of observations from dozens of monitoring stations, including rainfall, runoff, groundwater level, evaporation, and soil moisture), resulting in a correspondingly large adjacency matrix. To improve computational efficiency, a block-based sparse storage format can be adopted for the adjacency matrix, and parallel gradient calculation can be used to accelerate the iteration process. Furthermore, for variable pairs in the hydrological system with known physical laws (e.g., the unidirectional influence of upstream stations on downstream stations), directional constraints can be applied to corresponding elements of the adjacency matrix during the initialization phase. This incorporates known causal direction information into the optimization framework, further improving the accuracy and physical interpretability of causal structure learning.

[0040] In one optional implementation, based on the causal structure representation, reverse causal path tracing and multi-level propagation analysis are performed on the detected anomalous observed variables to identify all upstream root cause nodes. The causal contribution of each root cause node to the anomaly is quantified through intervention effect estimation and counterfactual comparison, forming a root cause node set and its corresponding contribution distribution, including: Based on the endpoint matching rules of directed edges, all incoming edges with the endpoint of the abnormal observation variable are filtered from the causal structure representation, the association is established, and the starting node of the incoming edge and the lag step number marked on the edge are recorded. For the starting node of the incoming edge, continue to parse its incoming edge, backtrack layer by layer along the direction of the incoming edge and mark the level depth until the node no longer has an incoming edge, and determine the node without an incoming edge as the upstream root cause node based on the feature of missing incoming edges. A directed path from each upstream root cause node to the abnormal observation variable is generated by depth-first traversal. The lag steps of each edge on the directed path are superimposed and transferred along the path direction to obtain the end-to-end time delay. The nonlinear mapping rules of each edge on the directed path are combined in the connection order to obtain the path cascade transfer function. For each upstream root cause node, the observed value of the node is reconstructed into the historical normal median value based on the historical distribution characteristics to construct the intervention state. The propagation time offset of the intervention state is inferred based on the end-to-end time delay corresponding to the node. The response value of the intervention state propagating through each directed path is calculated using the path concatenation transfer function, and the response values ​​are fused according to the propagation intensity of each directed path to obtain the counterfactual estimate of the abnormal observation variable. Calculate the deviation between the actual observed value and the counterfactual estimated value of the abnormal observed variable, and map the deviation to the causal contribution of the corresponding upstream root cause node through a difference metric; Sort and organize all upstream root cause nodes and their corresponding causal contributions according to the causal contribution degree, forming a root cause node set and its corresponding contribution degree distribution.

[0041] After obtaining the causal structure representation containing nodes, directed edges, and causal strength weights, reverse causal path tracing is performed for the detected anomalous observation variables. All directed edges in the causal structure representation are traversed, and all incoming edges ending with the anomalous observation variable are selected according to the endpoint matching rule. The starting node identifier and the lag step count marked on each incoming edge are recorded. The purpose of this step is to establish the association mapping between the anomalous observation variable and its direct predecessor node, providing a starting point for subsequent layer-by-layer backtracking. Recording the lag step count is crucial because the impact of upstream rainfall, runoff, and other variables on downstream water levels and flows in hydrological systems often has a physical time delay, which is reflected in the causal structure representation as lag step markings on the edges.

[0042] After filtering direct incoming edges, the incoming edges of each starting node are further parsed, backtracking upwards layer by layer along the directed edge direction. Simultaneously, the layer depth of each visited node relative to the outlier observed variable is marked. The layer depth starts from 1, with the depth of the direct predecessor node of the outlier observed variable being 1, the predecessor depth of the predecessor node being 2, and so on. The backtracking process continues until a node no longer has any incoming edges in the entire causal structure representation. Based on the missing incoming edge characteristic, such nodes are identified as upstream root cause nodes. In hydrological telemetry scenarios, upstream root cause nodes typically correspond to exogenous input variables such as rainfall, groundwater recharge, and upstream reservoir discharge. These variables are at the very top of the causal chain and are not causally driven by other variables.

[0043] After identifying all upstream root cause nodes, a depth-first search strategy is employed. Starting from each root cause node, the search proceeds downwards along the directed edges to generate all directed paths from that root cause node to the anomalous observed variable. When the depth-first search encounters an anomalous observed variable node, it records the complete path and backtracks, continuing to search the remaining branches to enumerate all propagation paths. For each directed path, the lag steps labeled on each edge are sequentially added along the path direction to obtain the end-to-end time delay L for that path. k , where k is the path number. The end-to-end time delay reflects the total number of time steps required for the state change of the root cause node to propagate to the anomalous observed variable via this path. Simultaneously, the nonlinear mapping laws corresponding to each edge of the path are composited according to the connection order to obtain the path cascade transfer function G. k The path cascade transfer function characterizes all the nonlinear transformations a signal undergoes as it propagates along the path, and is the core computational tool for subsequent counterfactual estimation.

[0044] For each upstream root cause node, based on its distribution characteristics in the historical time series, the historical normal median value is extracted as the intervention state. The historical normal median value is calculated with the time point of the anomalous event as the boundary, taking the median of observations within a sufficiently long window before that time point to avoid the interference of extreme values ​​on the reference benchmark. Replacing the actual observation value of the node with the historical normal median value constructs the intervention state for that node, simulating the counterfactual assumption that "the system should present itself if the node is at a normal level".

[0045] Based on the end-to-end time delay L corresponding to each directed path from the root cause node to the anomalous observed variable. k This involves performing a time-shifted projection of the intervention state. Specifically, the timestamp of the intervention state is shifted backward by L. k The number of steps is adjusted to align with the observation times of the outlier variables, ensuring temporal consistency of causal propagation. After time offsetting, the intervention state is substituted into the path cascade transfer function G. k Perform forward computation to obtain the response value r at the abnormal observation variable after the intervention state propagates through the k-th directed path. k When multiple directed paths exist, the response values ​​of each path are weighted and fused based on their propagation strength. The propagation strength is obtained by multiplying the causal strength weights of each edge on the path, denoted as s. k The fused counterfactual estimate It is obtained by summing and normalizing the weighted response values ​​of all paths, i.e. This counterfactual estimate represents the expected value that the anomalous observed variable should take under the assumption that the root cause node is in a normal state.

[0046] Calculate the actual observed value y of the outlier variable. obs Counterfactual estimates Deviation Δ between m Where m is the root cause node number. The deviation magnitude is calculated using the absolute difference form, i.e. ,in This represents the counterfactual estimate of the intervention state constructed for the m-th root cause node. A larger deviation indicates a more significant reduction in the bias of the observed variable if the root cause node recovers, meaning a stronger causal contribution of that node to the current anomaly. The deviation magnitude is mapped to causal contribution C using a difference metric. m The mapping method involves normalizing the deviation magnitude of all root cause nodes, i.e. This makes the sum of the causal contributions of all root cause nodes equal to 1, forming a probability distribution form that can be directly compared.

[0047] According to causal contribution C mAll upstream root cause nodes are sorted in descending order of magnitude, and the sorting results are organized with their corresponding contribution values ​​into a root cause node set and contribution distribution. Root cause nodes ranked higher indicate their strongest explanatory power for the current hydrological anomaly and are the priority targets for intervention. In practical hydrological control scenarios, such as when a monitoring section experiences an abnormal water level exceeding the warning level, the above process can quantitatively identify the contribution of factors such as a sudden increase in upstream tributary inflow, abnormal reservoir operation, or localized torrential rainfall, thus providing targeted decision-making basis for the generation of subsequent intervention and control plans. The output format of the contribution distribution allows operators to intuitively judge the primary and secondary relationships of the anomaly's causes and rationally allocate control resources in scenarios with multiple concurrent root causes.

[0048] In one optional implementation, for each upstream root cause node, the observed values ​​of the upstream root cause node are reconstructed into historical normal median values ​​based on historical distribution characteristics to construct an intervention state. The propagation time offset of the intervention state is deduced based on the end-to-end time delay corresponding to the upstream root cause node, including: For each upstream root cause node, the observation sequence of the upstream root cause node in the normal operation interval is extracted from the historical time window, and the historical normal median value is obtained by performing quantile statistics on the observation sequence in chronological order. Replace the actual observed value of the upstream root cause node at the time of the anomaly with the historical normal median value, and construct the intervention state observation value of the upstream root cause node as a counterfactual benchmark. Extract the end-to-end time delay on the directed path from the upstream root cause node to the abnormal observation variable, and use the end-to-end time delay as the propagation delay step number of the intervention state observation value from the upstream root cause node to the abnormal observation variable. Based on the propagation delay steps, the effective time of the intervention state observation is shifted backward by the time length corresponding to the propagation delay steps to obtain the arrival time of the intervention state at the abnormal observation variable. The time difference between the arrival time and the time of the abnormal occurrence of the abnormal observation variable is determined as the propagation time offset of the intervention state.

[0049] For each upstream root cause node, its normal operating state needs to be reconstructed from historical data to build a reliable counterfactual baseline. Specifically, for a given upstream root cause node, the observation sequence of that node is extracted from a sufficiently long historical time window, tracing back from the current anomaly occurrence. The length of this historical time window needs to cover a sufficient number of normal operating cycles to ensure the representativeness and stability of the statistical results. After extracting the historical observation sequence, it is necessary to further filter the observations that belong to the normal operating range and exclude observations that have experienced anomalies within the historical period, thereby ensuring that the data used for statistics comes from the actual records when the system was in a healthy operating state.

[0050] The selected normal operation observation sequence was statistically analyzed chronologically, and the 50th quantile of the sequence was calculated, which is the historical normal median. The median, rather than the mean, was chosen as the representative statistic for the normal state because it has inherent robustness to extreme outliers. In hydrological telemetry scenarios, the historical distribution of physical quantities such as water level, flow rate, and rainfall often exhibits long-tail characteristics or seasonal skewness. If the mean is used as a benchmark, it is easily affected by historical extreme events, causing the counterfactual benchmark to deviate from the true normal operation center. The median can more stably reflect the typical value level of a node under normal conditions, thus providing a more reliable reference for estimating subsequent intervention effects. Let the observation sequence of the upstream root cause node v within the historical normal interval be... , where N v The historical median value represents the number of normally functioning samples included in the statistics. Satisfy: When N v When it is an odd number, For the sorted number N elements; when N v When it is even, For the sorted number With the The average of the elements.

[0051] When constructing the intervention state, the actual observed value of the upstream root cause node v at the time of the anomaly occurrence is replaced with the aforementioned historical normal median value. The observed values ​​of the remaining variables remain unchanged, thus forming the intervention state observation value for this node. This substitution operation corresponds to the intervention operation in the "do-calculus" within the causal inference framework, which forcibly fixes the value of node v to the normal level, cutting off the influence of its upstream variables, thereby simulating a counterfactual scenario of "what state the system would be in if this node did not experience an anomaly." The intervention state observation value is denoted as... ,but This value serves as the starting point for subsequent counterfactual propagation calculations.

[0052] After determining the intervention state observations, it is necessary to quantify the time delay required for the intervention state to propagate from the upstream root node v to the anomalous observed variable. This involves extracting all directed paths from the upstream root node v to the anomalous observed variable from the causal graph model, and determining the end-to-end time delay L for each directed path. k This has been recorded in the causal structure representation. In the single-path case, the end-to-end time delay of that path is directly taken as the propagation delay steps. In the multi-path case, the end-to-end time delay corresponding to the path with the shortest propagation delay among all directed paths is taken as the delay steps of the main propagation channel, denoted as... This value represents the time step in which the intervention status signal can reach the abnormal observation variable. This approach aligns with the physical law in hydrological systems that signals propagate along the shortest path first; for example, runoff changes caused by upstream rainfall events will first affect downstream water level monitoring stations along the nearest confluence channel.

[0053] Based on the number of propagation delay steps , intervention status observation values The effective time is shifted backward. The time length corresponding to each time step. Let t be the time when the anomaly occurs in the observed variable. anom The moment when the intervention state of the upstream root node v takes effect is the moment when the anomaly occurs (t). anom Since the intervention replaces the value of node v at the moment the anomaly occurs, the arrival time of the intervention state at the anomaly observed variable is... satisfy: , where Δt is the system's sampling time step, i.e., the time interval between two adjacent telemetry data acquisitions.

[0054] Time of arrival The time t of the anomaly occurrence of the observed variable anom The time difference between them is determined as the propagation time offset of the intervention state, denoted as . ,but: This propagation time offset has a clear physical meaning: it describes how long it takes for a state change occurring at an upstream root node to have an observable impact on anomaly variables downstream. In hydrological control scenarios, this time offset corresponds to the runoff time from upstream monitoring points to downstream monitoring points, or the response lag time of rainfall signals transmitted to the river cross-section through hydrological processes such as surface runoff and interflow.

[0055] Propagation time offset It plays a crucial role in subsequent counterfactual effect estimation, when This indicates that the intervention state had not yet reached the abnormal observed variable at the time of the anomaly, and alignment correction needs to be performed on the time axis to ensure the counterfactual estimate. The calculations are performed under correct time alignment conditions, rather than simply comparing the immediate effects of the intervention state with the observations at the moment of the anomaly. Ignoring the propagation time offset would lead to an imbalance in the causal contribution C. m The estimation of propagation time offset introduces systematic biases, which in turn affect the quality of the generated intervention and control schemes. By introducing precise extrapolation of propagation time offset, the intervention effect and abnormal response can be accurately aligned in the time dimension, making the estimation of the causal contribution of each root cause node more accurate, and ultimately supporting the generation of more targeted and timely intelligent hydrological control schemes.

[0056] For different upstream root cause nodes, due to the different topological distances and path delay characteristics between them and the anomalous observed variables in the causal graph, the propagation time offsets for each node vary. There are typically differences. Root cause nodes that are topologically closer to the observed anomaly variable and have shorter path delays have smaller propagation time offsets, allowing intervention effects to manifest more quickly at the observed anomaly variable. Conversely, root cause nodes that are farther away and have larger path delays have larger propagation time offsets, requiring more time to produce observable effects downstream. This differentiated time offset information provides important reference for hydrological management decisions, enabling control personnel to rationally prioritize and sequence intervention operations based on the response time of different root cause nodes, thereby achieving optimal hydrological anomaly control effects under time constraints.

[0057] In one optional implementation, based on the root cause node set and the causal contribution distribution, combined with hydrological control constraints and system physical laws, an intervention control scheme is generated, including: Identify and index each upstream root cause node from the root cause node set and associate it with its corresponding causal contribution value in the causal contribution distribution, and construct an intervention priority ranking of the root cause nodes according to the causal contribution value. For each upstream root cause node in the intervention priority ranking, the hydrological control variables associated with the upstream root cause node and the physical adjustable range of the hydrological control variables are identified and established from the causal structure representation. The physical adjustable range is marked and constructed as the initial control space of the upstream root cause node. Obtain the hydrological control constraints of the hydrological system, and decompose and reorganize the hydrological control constraints into the constraint boundaries of the initial control space of each upstream root node; Based on the constraint boundary, the initial control space of each upstream root cause node is iteratively shrunk and adaptively adjusted to generate a feasible control space for each upstream root cause node. Within the feasible control space of the upstream root cause node, the control adjustment value is optimized and locked. The path cascade transfer function from the upstream root cause node to the abnormal observation variable and the end-to-end time delay are used to calculate the expected correction effect of the control adjustment value on the abnormal observation variable through the causal path. The control adjustment values ​​and expected correction effects of each upstream root cause node are sequentially arranged and aggregated into an intervention control scheme according to the intervention priority.

[0058] like Figure 2 As shown, the method includes: After identifying the root cause node set and the causal contribution distribution, the process of generating an intervention and control plan begins with establishing an intervention priority ranking for the root cause nodes. The root cause node set is traversed, and an index association is established for each upstream root cause node, binding it to the corresponding contribution value in the causal contribution distribution. Let the causal contribution value of the u-th upstream root cause node be C. u Press C u All upstream root cause nodes are sorted in descending order to form an intervention priority sequence. Where R represents the total number of root cause nodes, and u1 corresponds to the node with the highest contribution and the greatest need for priority intervention. The physical meaning of this ranking is that the higher the causal contribution of a root cause node, the more significant its current deviation has on the abnormal observed variable; prioritizing its control and regulation can achieve the greatest corrective benefit with minimal intervention cost. In practical hydrological scenarios, such as in a basin flood peak warning scenario, when multiple upstream hydrological stations simultaneously exhibit anomalies, priority ranking can help dispatchers quickly identify the most critical control nodes and avoid resource dispersion.

[0059] For each upstream root cause node in the intervention priority sequence, the hydrological control variables directly associated with that node are retrieved from the causal structure representation. Hydrological control variables refer to physical quantities that can be adjusted through operational means in actual engineering, such as gate opening, pump station flow, and reservoir discharge. For the u-th upstream root cause node, the set of associated control variables is identified, and the physical adjustable range of each control variable is read from the hydrological equipment parameter library or engineering specifications. Let the physical adjustable range of the q-th control variable associated with the u-th root cause node be denoted as . ,in This is the physical lower bound of the control variable. The upper physical bound is defined by combining the physically adjustable intervals of all control variables to construct the initial control space of the u-th upstream root cause node. This space is the Cartesian product of the physically adjustable intervals of each control variable, representing all theoretically achievable control states of the node without considering any external constraints.

[0060] After obtaining the hydrological control constraints of the hydrological system, they need to be decomposed and reorganized into constraint boundaries of the initial control space of each upstream root node. Hydrological control constraints typically include multiple levels: safety constraints, ecological constraints, engineering constraints, and multi-node linkage constraints. These constraints are decomposed according to the control variables of each root node, extracting the subset of constraints directly related to the control variables of the u-th root node, and transforming them into boundary restrictions in each dimension of the node's initial control space. For coupled constraints involving multi-node linkage, inequalities or equations are added to the control space boundaries of the relevant nodes to ensure that the control adjustment values ​​of each node meet the requirements of global water conservation and safe operation during subsequent optimization.

[0061] Based on the constraint boundaries, the initial control space of each upstream root cause node is iteratively shrunk and adaptively adjusted to generate a feasible control space. The core logic of the iterative shrinking process is: mapping each constraint boundary to each dimension of the initial control space, pruning regions exceeding the constraint range, and simultaneously checking whether the shrunken control space satisfies the intersection of all constraint conditions. If a constraint causes a conflict between the upper and lower bounds of the control space in a certain dimension (i.e., ... If the constraints are met after they are applied, an adaptive adjustment mechanism is triggered: lower-priority non-safety constraints are relaxed first (e.g., the lower limit of ecological flow can be appropriately reduced under extreme conditions), or some adjustment tasks are transferred to adjacent priority nodes, until the feasible control space is not empty. The iteration termination condition for adaptive adjustment is: all constraints are satisfied, and the interval width of the feasible control space no longer changes in any dimension, i.e., the contraction process converges. The final result is... It is the set of control states that can actually be executed by the u-th root cause node under the combined effect of the physically adjustable range and all hydrological control constraints.

[0062] Within feasible control space Internal optimization to determine the control adjustment value d u The optimization objective is to select the control adjustment value that maximizes the expected correction effect on the outlier observed variables, while satisfying all constraints. This is achieved using the path cascade transfer function G, which has been calibrated in the causal structure representation. k With end-to-end time delay L k The candidate control regulation value is substituted into the path propagation model to calculate the expected response at the outlier observed variable after the regulation value propagates through the causal path. Let the control regulation value of the u-th root cause node be d. u The expected correction contribution at the outlier observed variable after propagation via the k-th directed path is G. k (d) u By combining all directed paths originating from the u-th root cause node and reaching the abnormal observed variable, and weighting and fusing the contributions of each path, the expected correction effect E corresponding to the control adjustment value of that node is obtained. u The optimization process involves traversing candidate values ​​within the feasible control space using either a grid search or gradient projection method, selecting the value that makes E... u The largest d that satisfies all constraints u As a locked control adjustment value. End-to-end time delay L k The time delay between the issuance of the control adjustment value and the generation of its effect is determined. When arranging intervention plans, this time delay must be included in the calculation of the execution sequence to ensure that the adjustment actions of each node are reasonably arranged on the time axis and to avoid the mutual cancellation of the intervention effects of multiple paths due to time delay differences.

[0063] According to intervention priority sequence The control adjustment value d of each upstream root cause node u The expected correction effect E u The intervention and control schemes are sequentially arranged and aggregated to form a complete plan. This structured plan includes the following elements: the execution order of each root cause node, the target adjustment value of each control variable, the expected effective time of each node's adjustment action (determined jointly by the anomaly occurrence time and end-to-end time delay), and the expected correction trajectory of the observed abnormal variables after intervention at each node. For nodes with high causal contribution, their adjustment actions are prioritized and executed at the top of the plan. For nodes with lower contribution but with interconnected constraints, their execution sequence must coordinate with higher-priority nodes to ensure that the overall water balance is not disrupted. The final output intervention and control scheme can be directly transmitted to the hydrological scheduling execution layer, driving gates, pumping stations, and other actuators to complete their adjustment actions according to the predetermined sequence, achieving effective intervention and correction of abnormal hydrological conditions.

[0064] In one optional implementation, the structural topology and causal strength weights of the causal graph model are updated online using the feedback observation data, and parameter correction is performed to address the deviation between the actual intervention effect and the predicted effect, including: The time-series observation values ​​of each monitoring node after the intervention are analyzed from the feedback observation data. The time-series observation values ​​are then aligned with the historical observation values ​​of the corresponding nodes before the intervention, and the change in the observation values ​​of each node before and after the intervention is calculated. The observed changes are compared node by node with the expected correction effect preset in the intervention and control plan, and the deviation between the actual intervention effect and the expected correction effect is calculated. Based on the deviation value, identify the source of prediction error of the propagation path in the causal graph model. For causal edges where the prediction error exceeds the allowable range, extract the node pairs connected by the causal edge and the causal strength weight labeled on the causal edge. Based on the actual response relationship of node pairs in the feedback observation data, the causal strength weight is numerically corrected, and the corrected causal strength weight is written back to the corresponding causal edge of the causal graph model. Detect whether there are any new causal relationships in the feedback observation data that are not covered by the causal graph model. If there are new causal relationships, insert new causal edges into the structural topology of the causal graph model and label the initial causal strength weights. The updated causal strength weights and structural topology are synchronized to the causal graph model to complete the online update and parameter correction of the causal graph model.

[0065] After the intervention and control plan is implemented, the time-series observation sequence after the intervention is extracted from the feedback observation data continuously collected by each monitoring node. Due to the sampling frequency, transmission delay, and clock deviation of different monitoring nodes, there are misalignments in the data on the time axis. It is necessary to perform time-series alignment processing on the time-series observations after the intervention of each node with the historical observations of the corresponding node before the intervention. The time-series alignment uses the time of the intervention command as the reference anchor point, and maps the observation sequences before and after the intervention to a unified time coordinate axis according to the actual sampling time step of each node. For nodes with inconsistent sampling intervals, linear interpolation is used to fill in the missing observations at the time, ensuring that the observations of the same node before and after the intervention correspond one-to-one in the time dimension. After the time-series alignment is completed, the difference between the post-intervention observation and the historical observation at the corresponding time before the intervention is calculated for each monitoring node at each time step, and the change sequence of the observation values ​​of each node is obtained. This change sequence reflects the actual response amplitude and response timing caused by the intervention operation at each monitoring location in the hydrological system.

[0066] The changes in observed values ​​at each node are compared with the expected correction effect E preset for the corresponding node in the intervention and control plan. u Perform a node-by-node comparison. For each monitoring node associated with the u-th root cause node, calculate the change in the actual observed value and E. u Deviation between ,Right now , where Δ obs，u This represents the actual change in the observed value after intervention at that node. When the deviation exceeds a pre-set tolerance threshold θ, the causal propagation path corresponding to that node is determined to have a prediction error, requiring subsequent causal edge weight correction. The positive and negative directions of the deviation value also have physical meaning: a positive deviation indicates that the actual intervention response is stronger than expected, suggesting that the corresponding causal strength weight is underestimated in the causal graph model; a negative deviation indicates that the actual response is weaker than expected, suggesting that the corresponding causal strength weight is overestimated. By statistically summarizing the deviation values ​​of all monitoring nodes, specific propagation path segments with insufficient prediction accuracy in the causal graph model can be located.

[0067] Regarding the deviation value For causal edges exceeding the allowable threshold θ, extract the node pairs connected to the causal edge and their currently labeled causal strength weights W from the structural representation of the causal graph model. pq Here, the subscripts p and q correspond to the source and target node numbers of the causal edge, respectively. Based on the actual response relationship of node pairs (p, q) in the feedback observation data, a local regression fitting method is used to re-estimate the actual influence coefficient of the change in the observed value of node p on the change in the observed value of node q, thus obtaining the corrected causal strength weight estimate. To avoid drastic weight fluctuations caused by noise interference in a single feedback data session, an exponentially weighted smoothing mechanism is introduced to progressively update the weights. The update rule is as follows: α is the weight update step size coefficient, ranging from (0, 1), which controls the fusion ratio between historical weights and new estimates. The value of α should be configured in conjunction with the dynamic change rate of the hydrological system. During seasons when the system state is relatively stable, α can be appropriately reduced to maintain model stability, while during flood season or sudden hydrological events, α can be appropriately increased to improve the model's response speed to environmental changes. After numerical correction, [the following will be done]. Write back to the weight attribute of the corresponding causal edge in the causal graph model to complete the parameter correction of the causal edge.

[0068] After correcting the weights of existing causal edges, the system further detects whether there are any new causal relationships in the feedback observation data that are not covered by the current causal graph model structure. The detection method involves calculating the conditional mutual information of the temporal change sequences of all node pairs in the feedback observation data. For node pairs that are not connected by directed edges in the current causal graph model... If its conditional mutual information value exceeds the threshold for introducing a new edge If the correlation between nodes remains statistically significant after excluding known common causal variables, then a new causal relationship is determined to exist between the node pairs. For the identified new causal relationship, its directionality needs further verification: by comparing nodes... Change sequences and nodes Based on the chronological order of the changes on the timeline, and combined with prior knowledge of the physical propagation direction of the hydrological system, the direction of the causal edge is determined as follows. or After confirming the direction, the corresponding new directed causal edge is inserted into the structural topology of the causal graph model, and the initial causal strength weight of the new causal edge is estimated based on the actual response strength of the node pairs in the feedback observation data. The initial weights are also estimated using a local regression method to intervene in the preceding and following nodes. With nodes The ratio of the changes in observed values ​​is used as the basic estimate of the initial weights, and the propagation attenuation effect is appropriately corrected by combining the path length.

[0069] After inserting a new causal edge, the structural validity of the entire causal graph model needs to be verified to ensure that the graph structure still satisfies the directed acyclic graph constraint. If the added directed edge causes a cycle in the graph, edges with smaller weights or weaker physical meanings are selectively deleted based on the weights of the new edge and existing edges, as well as their physical rationality, to eliminate the cycle and maintain the directed acyclic property of the causal graph. After completing the structural validity verification, all updated causal strength weights and corrected structural topology are synchronously written into the storage representation of the causal graph model, including updating the values ​​of corresponding elements in the weighted adjacency matrix, updating the edge attribute records of node pairs, and updating the model's version timestamp. After synchronization, the updated causal graph model takes effect immediately. Subsequent anomaly detection, root cause tracing, and intervention scheme generation are all performed based on the latest version of the causal graph model, thereby achieving the continuous adaptive updating and parameter correction capability of the causal graph model as the hydrological system dynamically evolves.

[0070] A second aspect of this invention provides an intelligent control system based on AI hydrological telemetry, comprising: The causal construction unit is used to acquire multi-source hydrological telemetry data, construct a causal graph model, and identify causal dependencies, time delay propagation characteristics and nonlinear coupling modes among variables from the multi-source hydrological telemetry data through a constrained optimization algorithm, thereby obtaining a causal structure representation containing nodes, directed edges and causal strength weights. The root cause analysis unit is used to perform reverse causal path tracing and multi-level propagation analysis on the detected abnormal observation variables based on the causal structure representation, identify all upstream root cause nodes, and quantify the causal contribution of each root cause node to the anomaly through intervention effect estimation and counterfactual comparison, forming a root cause node set and corresponding contribution distribution. An intervention generation unit is used to generate an intervention control scheme based on the root cause node set and the causal contribution distribution, combined with hydrological control constraints and system physical laws. The feedback update unit is used to execute the intervention control scheme and collect feedback observation data after execution. It uses the feedback observation data to update the structural topology and causal strength weights of the causal graph model online, and at the same time corrects the deviation between the actual intervention effect and the predicted effect.

[0071] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0072] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0073] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent control method based on AI hydrological telemetry, characterized in that, include: Acquire multi-source hydrological telemetry data, construct a causal graph model, and use a constrained optimization algorithm to identify causal dependencies, time delay propagation characteristics, and nonlinear coupling modes among variables in the multi-source hydrological telemetry data, thereby obtaining a causal structure representation that includes nodes, directed edges, and causal strength weights; Based on the aforementioned causal structure representation, reverse causal path tracing and multi-level propagation analysis are performed on the detected anomalous observed variables to identify all upstream root cause nodes. The causal contribution of each root cause node to the anomaly is quantified through intervention effect estimation and counterfactual comparison, forming a set of root cause nodes and their corresponding contribution distributions, including: Based on the endpoint matching rules of directed edges, all incoming edges with the endpoint of the abnormal observation variable are filtered from the causal structure representation, the association is established, and the starting node of the incoming edge and the lag step number marked on the edge are recorded. For the starting node of the incoming edge, continue to parse its incoming edge, backtrack layer by layer along the direction of the incoming edge and mark the level depth until the node no longer has an incoming edge, and determine the node without an incoming edge as the upstream root cause node based on the feature of missing incoming edges. A directed path from each upstream root cause node to the abnormal observation variable is generated by depth-first traversal. The lag steps of each edge on the directed path are superimposed and transferred along the path direction to obtain the end-to-end time delay. The nonlinear mapping rules of each edge on the directed path are combined in the connection order to obtain the path cascade transfer function. For each upstream root cause node, the observed value of the node is reconstructed into the historical normal median value based on the historical distribution characteristics to construct the intervention state. The propagation time offset of the intervention state is inferred based on the end-to-end time delay corresponding to the node. The response value of the intervention state propagating through each directed path is calculated using the path concatenation transfer function, and the response values ​​are fused according to the propagation intensity of each directed path to obtain the counterfactual estimate of the abnormal observation variable. Calculate the deviation between the actual observed value and the counterfactual estimated value of the abnormal observed variable, and map the deviation to the causal contribution of the corresponding upstream root cause node through a difference metric; Sort and organize all the upstream root cause nodes and their corresponding causal contribution values ​​according to the causal contribution values ​​to form a root cause node set and a corresponding contribution distribution. Based on the root cause node set and the causal contribution distribution, combined with hydrological control constraints and system physical laws, an intervention and control scheme is generated. The intervention and control scheme is executed and feedback observation data is collected after execution. The feedback observation data is used to update the structural topology and causal strength weights of the causal graph model online, and the parameters are corrected for the deviation between the actual intervention effect and the predicted effect.

2. The method according to claim 1, characterized in that, Multi-source hydrological telemetry data are acquired, a causal graph model is constructed, and a constrained optimization algorithm is used to identify causal dependencies, time-delay propagation characteristics, and nonlinear coupling modes among variables from the multi-source hydrological telemetry data, resulting in a causal structure representation including: Mutual information calculation and conditional independence test are performed on multi-source hydrological telemetry data to quantify the statistical correlation strength and conditional dependence between any two variables. Based on the statistical correlation strength and conditional dependence, gradient iteration is performed by applying acyclic constraints and sparsity penalties to obtain an initial directed acyclic graph. For each pair of variables corresponding to a directed edge in the initial directed acyclic graph, cross-correlation analysis is used to search for the number of lag steps that maximize the causal transmission effect within multiple candidate time lag intervals. The number of lag steps and the strength of the causal transmission effect are then labeled on the corresponding directed edges to obtain a time lag labeling graph containing time lag information. For each directed edge in the time-delay annotation graph, the historical observations of the source node at the corresponding lag step and the current observations of the target node are extracted to form a sample pair. The mapping pattern from the historical observations of the source node to the current observations of the target node is fitted by a nonlinear regression method to characterize the nonlinear coupling mode between variables. The topological structure of the time-delay annotation map, the number of lag steps, and the fitted nonlinear mapping pattern are integrated into a causal structure representation, which simultaneously encodes causal dependencies, time-delay propagation characteristics, and nonlinear coupling modes.

3. The method according to claim 2, characterized in that, Based on the statistical correlation strength and the conditional dependency, gradient iteration is performed by applying acyclic constraints and sparsity penalties to obtain the initial directed acyclic graph, including: An adjacency matrix is ​​constructed to represent the directed connections between variables, and the elements of the adjacency matrix are initialized and assigned values ​​according to the statistical association strength. The trace of the matrix exponential function is calculated based on the adjacency matrix to form a loop quantization index. The cumulative amount of the absolute value of the elements of the adjacency matrix and the loop quantization index are respectively set as the acyclicity constraint term and the sparsity penalty term. A joint loss function is constructed by combining the fitting error of the conditional dependence degree. The gradient of the joint loss function is calculated for the elements of the adjacency matrix. The element values ​​of the adjacency matrix are iteratively corrected according to the opposite direction of the gradient. During the correction process, the weight coefficient of the sparsity penalty term is adaptively determined according to the descent rate of the loop quantization index. In the iterative stage where the loop quantization index decreases slowly, the weight of the sparsity penalty term is increased to accelerate the edge pruning. Repeatedly calculate the loop quantization index, the gradient, and the adjacency matrix correction until the loop quantization index converges to the zero neighborhood and the joint loss function no longer decreases. Then, map the final adjacency matrix to generate the initial directed acyclic graph.

4. The method according to claim 1, characterized in that, For each upstream root cause node, the observed values ​​of the upstream root cause node are reconstructed into historical normal median values ​​based on historical distribution characteristics to construct the intervention state. The propagation time offset of the intervention state is inferred based on the end-to-end time delay corresponding to the upstream root cause node, including: For each upstream root cause node, the observation sequence of the upstream root cause node in the normal operation interval is extracted from the historical time window, and the historical normal median value is obtained by performing quantile statistics on the observation sequence in chronological order. Replace the actual observed value of the upstream root cause node at the time of the anomaly with the historical normal median value, and construct the intervention state observation value of the upstream root cause node as a counterfactual benchmark. Extract the end-to-end time delay on the directed path from the upstream root cause node to the abnormal observation variable, and use the end-to-end time delay as the propagation delay step number of the intervention state observation value from the upstream root cause node to the abnormal observation variable. Based on the propagation delay steps, the effective time of the intervention state observation is shifted backward by the time length corresponding to the propagation delay steps to obtain the arrival time of the intervention state at the abnormal observation variable. The time difference between the arrival time and the time of the abnormal occurrence of the abnormal observation variable is determined as the propagation time offset of the intervention state.

5. The method according to claim 1, characterized in that, Based on the root cause node set and the causal contribution distribution, combined with hydrological control constraints and system physical laws, the following intervention and control schemes are generated: Identify and index each upstream root cause node from the root cause node set and associate it with its corresponding causal contribution value in the causal contribution distribution, and construct an intervention priority ranking of the root cause nodes according to the causal contribution value. For each upstream root cause node in the intervention priority ranking, the hydrological control variables associated with the upstream root cause node and the physical adjustable range of the hydrological control variables are identified and established from the causal structure representation. The physical adjustable range is marked and constructed as the initial control space of the upstream root cause node. Obtain the hydrological control constraints of the hydrological system, and decompose and reorganize the hydrological control constraints into the constraint boundaries of the initial control space of each upstream root node; Based on the constraint boundary, the initial control space of each upstream root cause node is iteratively shrunk and adaptively adjusted to generate a feasible control space for each upstream root cause node. Within the feasible control space of the upstream root cause node, the control adjustment value is optimized and locked. The path cascade transfer function from the upstream root cause node to the abnormal observation variable and the end-to-end time delay are used to calculate the expected correction effect of the control adjustment value on the abnormal observation variable through the causal path. The control adjustment values ​​and expected correction effects of each upstream root cause node are sequentially arranged and aggregated into an intervention control scheme according to the intervention priority.

6. The method according to claim 1, characterized in that, The feedback observation data is used to update the structural topology and causal strength weights of the causal graph model online, and parameter correction is performed to address the deviation between the actual intervention effect and the predicted effect, including: The time-series observation values ​​of each monitoring node after the intervention are analyzed from the feedback observation data. The time-series observation values ​​are then aligned with the historical observation values ​​of the corresponding nodes before the intervention, and the change in the observation values ​​of each node before and after the intervention is calculated. The observed changes are compared node by node with the expected correction effect preset in the intervention and control plan, and the deviation between the actual intervention effect and the expected correction effect is calculated. Based on the deviation value, identify the source of prediction error of the propagation path in the causal graph model. For causal edges where the prediction error exceeds the allowable range, extract the node pairs connected by the causal edge and the causal strength weight labeled on the causal edge. Based on the actual response relationship of node pairs in the feedback observation data, the causal strength weight is numerically corrected, and the corrected causal strength weight is written back to the corresponding causal edge of the causal graph model. Detect whether there are any new causal relationships in the feedback observation data that are not covered by the causal graph model. If there are new causal relationships, insert new causal edges into the structural topology of the causal graph model and label the initial causal strength weights. The updated causal strength weights and structural topology are synchronized to the causal graph model to complete the online update and parameter correction of the causal graph model.

7. An intelligent control system based on AI hydrological telemetry, used to implement the method as described in any one of claims 1-6, characterized in that, include: The causal construction unit is used to acquire multi-source hydrological telemetry data, construct a causal graph model, and identify causal dependencies, time delay propagation characteristics and nonlinear coupling modes among variables from the multi-source hydrological telemetry data through a constrained optimization algorithm, thereby obtaining a causal structure representation containing nodes, directed edges and causal strength weights. The root cause analysis unit is used to perform reverse causal path tracing and multi-level propagation analysis on the detected abnormal observation variables based on the causal structure representation, identify all upstream root cause nodes, and quantify the causal contribution of each root cause node to the anomaly through intervention effect estimation and counterfactual comparison, forming a root cause node set and corresponding contribution distribution. An intervention generation unit is used to generate an intervention control scheme based on the root cause node set and the causal contribution distribution, combined with hydrological control constraints and system physical laws. The feedback update unit is used to execute the intervention control scheme and collect feedback observation data after execution. It uses the feedback observation data to update the structural topology and causal strength weights of the causal graph model online, and at the same time corrects the deviation between the actual intervention effect and the predicted effect.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images