Mountain torrent disaster dynamic plan generation method, product and equipment based on causal inference

Abstract

The invention discloses a method, a product and equipment for generating a mountain torrent disaster dynamic plan based on causal inference, and belongs to the technical field of mountain torrent disaster monitoring. The method comprises the following steps: S1, acquiring multi-source data, and extracting causal features of the multi-source data; during extraction, dual robust data deviation elimination is realized based on a tendency scoring model and a result regression model; s2, performing dynamic causal graph modeling based on the extracted causal features and a historical disaster data set, and updating the weight to obtain a causal graph with a time sequence weight; s3, determining a current disaster situation state according to the causal diagram information, and performing anti-fact plan generation in combination with a preset intervention measure set to obtain an optimal plan set; and S4, performing multi-department collaborative execution based on the optimal plan set. According to the method, the pain point problems of passive response, lack of causal support and cross-department collaboration low efficiency of a traditional mountain torrent plan can be solved.

Description

Technical Field

[0001] This invention belongs to the field of flash flood disaster monitoring technology, specifically relating to a method, product, and equipment for generating dynamic flash flood disaster emergency plans based on causal inference. Background Technology

[0002] Flash floods refer to floods, mudslides, and landslides caused by rainfall in hilly areas. In recent years, flash floods triggered by sudden, localized, and extreme heavy rainfall in my country have resulted in numerous casualties, accounting for an increasing proportion of total flood-related deaths, with mass casualty incidents occurring frequently. Flash floods are characterized by their suddenness, high destructiveness, and difficulty in early warning. To strengthen flash flood risk management and decision-making, and to guide flash flood prevention and control, it is essential to conduct flash flood monitoring and develop contingency plans.

[0003] Traditional emergency response plans are static. On the one hand, they are outdated, relying on historical disaster data and fixed rule bases, failing to integrate real-time hydrological, meteorological, and associated disaster causal relationships (such as landslides and bridge blockages), resulting in low alignment between the plan and the actual disaster situation. On the other hand, information on the causal coupling of data is lacking; the causal chain of rainfall-soil moisture content-runoff generation cannot be quantified, and plans are generated solely through statistical correlation, ignoring key causal mechanisms and significantly reducing their accuracy. Furthermore, their counterfactual reasoning capabilities are weak, failing to simulate the potential consequences of different intervention measures (such as reservoir scheduling and mass evacuation), leading to insufficient scientific rigor in the plans.

[0004] It is evident that the existing static contingency planning methods can no longer meet the high requirements of this field for flash flood contingency plans, both now and in the future. Improving the real-time nature, scientific rigor, and accuracy of flash flood contingency plans has become a key issue. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.

[0006] Therefore, the purpose of this invention is to provide a method, product and equipment for generating dynamic emergency plans for flash floods based on causal inference, which can solve the pain points of traditional flash flood emergency plans such as "passive response, lack of causal support and inefficient cross-departmental collaboration".

[0007] To solve the above-mentioned technical problems, the present invention is implemented as follows: This invention provides a method for generating dynamic emergency plans for flash floods based on causal inference. The method includes the following steps: S1. Obtain multi-source data and extract causal features from it; S2. Based on the extracted causal features and historical disaster datasets, a dynamic causal graph model is constructed and the weights are updated to obtain a causal graph with time-series weights. S3. Determine the current disaster status based on the cause-effect diagram information, and generate counterfactual contingency plans by combining the preset set of intervention measures to obtain the optimal contingency plan set; S4. Based on the optimal set of contingency plans, conduct multi-departmental collaborative execution.

[0008] In addition, the method for generating dynamic emergency plans for flash floods based on causal inference according to the present invention may also have the following additional technical features: In some of these implementations, the multi-source data includes meteorological radar data, BeiDou surface deformation monitoring data, reservoir scheduling logs, and traffic network status sensor data.

[0009] In some implementations, the causal feature extraction process in step S1 includes: Establish a spatiotemporal fusion matrix: calculate the terrain weight coefficient using the analytic hierarchy process and optimize the soil moisture sensitivity parameters using gradient descent. Data bias elimination: A robust bias elimination mechanism is achieved using both a propensity score model and an outcome regression model. The propensity score model is implemented using an XGBoost classifier, and the outcome regression model uses a random forest regression model. Output: The spatiotemporally aligned causal feature matrix is ​​obtained based on the determined causal relationships and then output.

[0010] In some implementations, step S2 includes: Determining the causal direction: Based on the additive noise model, if the variable Y Can be represented as a variable X The function and independent noise e The sum of, and noise e and X If statistical independence is established, then... X → Y Causal edge; Quantify causal chains: Use probability tree reasoning to determine whether the transmission probability of critical paths meets preset requirements, and activate an early warning when it does; Update causal weights: Calculate the deformation rate based on monitoring data, then calibrate the water level change rate parameter by combining it with historical dam break data, and dynamically update the weights based on the water level change rate parameter.

[0011] In some of these implementations, noise is determined. e and X The method for statistical independence is to verify it using the HSIC test. When the HSIC test shows that... p When the value is less than 0.01, the two are considered to be independent.

[0012] In some implementations, step S3 includes: Generative Adversarial Simulation: Adversarial simulation is performed using a generative adversarial network. The generator adopts a U-Net structure, and spectral normalization constraints are introduced into the discriminator. Evaluation of the effectiveness of the contingency plan: Causal distillation trees were used to evaluate the effectiveness of the contingency plan in order to maximize the difference in treatment effects between groups; Optimal solution selection: Pareto optimal solution selection is performed using the improved NSGA-II algorithm, with the objective of minimizing casualties and economic losses, and the constraint being the risk of dam failure.

[0013] In some implementations, step S4 includes: Two-layer reinforcement learning scheduling: the water conservancy layer uses a DQN network to update the flood discharge volume, and the emergency response layer performs transfer optimization based on the reward function; LBS fence trigger command: Achieve precise response based on spatial grid by dividing the space into Thiessen polygonal grids; Digital twin verification: Dynamic model correction based on real-time data is performed through a model correction mechanism.

[0014] In some implementations, the model calibration mechanism includes: determining the learning rate using gradient descent and adjusting model parameters using the Euclidean distance between the predicted and true values.

[0015] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method for generating dynamic emergency plans for flash flood disasters based on causal inference as described in any of the preceding embodiments.

[0016] This invention also provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for generating dynamic emergency plans for flash flood disasters based on causal inference as described in any of the preceding embodiments.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: In this embodiment of the invention, the method for generating dynamic emergency plans for flash flood disasters based on causal inference integrates the advantages of XGBoost and random forest through a dual robust algorithm. It eliminates data bias from two dimensions: "sample weight adjustment" and "result prediction correction", generating a spatiotemporally consistent and causally clear feature matrix. This solves the pain point of "lack of data causal coupling" in traditional methods and is the underlying data foundation of the entire dynamic emergency plan generation process. In this embodiment of the invention, the provided method for generating dynamic emergency plans for flash floods based on causal inference achieves dynamic simulation of the causal chain of flash floods through a "generation-discrimination" adversarial learning mechanism. The U-Net structure is suitable for processing spatiotemporally correlated data (such as the spatial distribution and temporal evolution of flash floods), and retains multi-scale features through skip connections, improving the detail accuracy of the generated results. The spectral normalization constraint can solve the problem of instability in traditional GAN ​​training, enabling the generator to more reliably simulate the causal effects of different intervention measures, providing a scientific basis for multi-objective optimization (such as minimizing casualties and economic losses). In this embodiment of the invention, the method for generating dynamic emergency plans for flash flood disasters based on causal inference provides a dual synergy of causal distillation tree assessment and NSGA-II to achieve structured trade-offs of causal effects, realizing a closed loop from "causal effect analysis" to "multi-objective decision-making", and solving the pain points of traditional plans that "lack scientific trade-offs and inefficient cross-departmental collaboration". In this embodiment of the invention, the method for generating dynamic emergency plans for flash floods based on causal inference has a breakthrough in dynamism: the response time for plan generation is shortened from hours to within 10 minutes, and minute-level causal graph updates are supported; In this embodiment of the invention, the method for generating dynamic emergency plans for flash flood disasters based on causal inference has achieved a scientific improvement: the accuracy of counterfactual inference has been increased to 92% (compared to 68% of the traditional method), and the false alarm rate has been reduced to below 3%. In this embodiment of the invention, the method for generating dynamic emergency plans for flash floods based on causal inference achieves optimized collaborative efficiency: the cross-departmental instruction synchronization delay is reduced from 15 minutes to 2 minutes; In this embodiment of the invention, the method for generating dynamic emergency plans for flash flood disasters based on causal inference has significant economic benefits: it can reduce the direct economic losses caused by the failure of emergency plans by approximately RMB 180 million per year.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for generating dynamic emergency plans for flash floods based on causal inference, as disclosed in an embodiment of the present invention. Detailed Implementation

[0020] 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, not all, of the embodiments of the present invention. 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.

[0021] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.

[0022] Please see Figure 1 As shown, in some embodiments of the present invention, a method for generating dynamic emergency plans for flash flood disasters based on causal inference is provided, constructing an architecture of "data fusion - causal modeling - plan deduction - collaborative execution". This method mainly includes: Dynamic causal modeling: Constructing a causal graph model of flash flood disasters and quantifying the causal weights of multiple factors (rainfall, topography, engineering scheduling); Counterfactual scenario simulation: simulating the potential outcomes of different intervention strategies based on generative adversarial networks; Cross-departmental intelligent collaboration: Breaking down data barriers between water conservancy, emergency response, and transportation to generate multi-objective optimized joint response plans.

[0023] In some embodiments of the present invention, the specific steps of the method for generating dynamic emergency plans for flash flood disasters based on causal inference include: Step 1: Multi-source data acquisition and causal feature extraction The objects of operation are the multi-source data collected, including meteorological radar data, BeiDou surface deformation monitoring data, reservoir scheduling logs, and traffic network status sensor data.

[0024] The process of causal feature extraction includes: Step 1.1: Establish the spatiotemporal fusion matrix

[0025] Among them, the terrain weight coefficient w i The soil moisture sensitivity parameter λ was calculated using the analytic hierarchy process (DEM resolution ≤ 30m) and optimized using gradient descent (learning rate η = 0.01).

[0026] M ( t ) represents the spatiotemporal fusion matrix (the final output, used to characterize the fusion weights or fusion results of data at different spatiotemporal scales); t Represents time variables (different time points or moments); n : Indicates the total number of spatial units participating in the fusion (such as the number of spatial units like pixels and regions); wi Indicates the first i The weighting coefficient of each spatial unit (an input parameter used to measure the importance of different spatial units in the fusion process; it needs to be preset or calculated by other methods). Indicates the first i A spatial unit in time t The observational data (such as spectral values ​​of remote sensing images, surface parameters, etc.) for spatial coordinates x i The partial derivative (input parameter, representing the spatial gradient of the observation data of this spatial cell); DEM i Indicates the first i The gradient of a digital elevation model (DEM) for each spatial cell (input parameter, representing the spatial rate of change of topographic features such as slope and aspect). l This represents the regularization parameter (an input parameter used to balance the contributions of the two terms in the formula, avoiding overfitting or highlighting the influence of a certain factor; it needs to be adjusted according to the actual data). Soil ( t () indicates time t Observed or simulated values ​​of soil-related parameters (such as soil moisture, soil reflectivity, etc.) (input parameters that reflect the spatiotemporal changes of soil properties); Ssat This represents the reference or baseline value from satellite observations (input parameter, typically a long-term average or standard value from high spatiotemporal resolution satellite data, used to normalize soil parameters).

[0027] The input-output relationship is as follows: Input: Time variable t Total number of spatial units n Weights of each spatial unit w i Spatial gradient of observation data DEM gradient DEM i Regularization parameters l Soil parameters Soil ( t ), satellite reference value Ssat Output: Spatiotemporal fusion matrix M ( t (This integrates the spatiotemporal information of spatial topographic features, observation data gradients, and soil parameters, and is used for subsequent multi-source data fusion or spatiotemporal interpolation.)

[0028] The approach to obtaining the formula is as follows: Spatial feature fusion: through This method combines the spatial gradient of observation data with the topographic gradient to capture the influence of topography on the spatial distribution of observation data (such as the impact of differences in illumination and temperature caused by topographic undulations on remote sensing observations).

[0029] Spatiotemporal weight allocation: through weights w i The contributions of different spatial units are weighted and summed to highlight the role of key areas (such as areas with complex terrain or data-sensitive areas).

[0030] Soil parameter regularization: Introduction This method utilizes the ratio (logarithmic form) between soil parameters and satellite baseline values ​​to constrain the spatiotemporal consistency of the fusion results, avoiding fusion deviations caused by local observation errors.

[0031] The problems to be solved include: Spatiotemporal mismatch of multi-source data: Integrating high-resolution topographic data (DEM) with low-resolution temporal observation data (such as soil parameters) to resolve the differences in spatiotemporal scales between data from different sensors; Topographic impact correction: Corrects the interference of topographic relief on observation data by using the DEM gradient term, thereby improving the fusion accuracy of mountainous and complex terrain areas; Data uncertainty control: By using the regularization parameter λ and the satellite reference value Ssat, observation noise is smoothed to ensure the stability and physical rationality of the fusion results.

[0032] Advantages include: Multi-factor synergy: Comprehensive consideration of spatial topography (DEM gradient), dynamic changes in observation data (spatial gradient), and surface properties (soil parameters) to achieve multi-dimensional information fusion.

[0033] Spatiotemporal adaptability: through the time variable t and weights w i It can flexibly adapt to the data characteristics of different time points and spatial regions, and is suitable for dynamic monitoring scenarios (such as agricultural drought monitoring and ecological environment assessment).

[0034] The physical meaning is clear: all items are based on geophysical processes (topography-observation data relationship, soil-satellite data correlation), avoiding the black box problem of pure mathematical models, and the results are highly interpretable.

[0035] Step 1.2: Apply a dual robust algorithm to eliminate data bias: In flash flood disaster data, the raw data may contain selection bias (such as missing data at some monitoring points due to complex terrain) or confounding factors (such as collinearity between rainfall data and soil moisture data), directly affecting the accuracy of causal inference. The dual robust algorithm combines a propensity score model and an outcome regression model to achieve dual control over data bias, ensuring the reliability of subsequent causal modeling.

[0036] The formula for eliminating data bias is:

[0037] : Dual robust estimator (the average treatment effect of the final estimate).

[0038] N Sample size.

[0039] i Individual index (i=1,2,...,N) i =1,2,..., N ).

[0040] T i ： Handling assignment variables ( T i =1 indicates an individual i Accept processing. T i =0 indicates that no processing was received.

[0041] Y i Outcome variable (individual) i (Observational results).

[0042] X i Covariate vector (individual) i The characteristic variables are used to control for confounding.

[0043] Propensity score estimate (individual) i The probability of being processed. .

[0044] : Estimated values ​​of the model for the treatment group results .

[0045] The estimated value of the model for the control group results .

[0046] Propensity rating model e ( XThe XGBoost classifier (AUC≥0.85) was used. Selection bias (such as bias caused by analyzing only data from easily monitored areas) was eliminated by adjusting sample weights.

[0047] Result regression model m ( X Random forest regression (R² ≥ 0.9) was used. By fitting the mapping relationship between features and results, measurement bias or model specification bias was eliminated.

[0048] The output is: a spatiotemporally aligned causal feature matrix. F ( t )∈ ^{ n × m}( n For the number of monitoring points, m (For feature dimensions).

[0049] Propensity score refers to a score based on a given characteristic. X Given factors such as rainfall, soil type, and terrain slope, this represents the probability that a sample will be assigned to a specific "treatment group" (e.g., an area prone to landslides). Here, "treatment" can be understood as potential disaster-causing factors, and the tendency scoring model... e ( X XGBoost is used to quantify the probability of various factors influencing the occurrence of disasters. It is a high-efficiency gradient boosting tree algorithm, adept at handling high-dimensional features and non-linear relationships, and suitable for classification problems (such as determining whether a landslide will occur in a certain area). It requires the model's AUC (area under the curve) to be ≥0.85. AUC is a metric for classifier performance, ranging from 0 to 1. A value of 0.85 indicates that the model can effectively distinguish between the "disaster-occurred group" and the "non-disaster-occurred group," with a low false positive rate.

[0050] Propensity score models can identify key features that cause data bias (such as incomplete data collection in areas with complex terrain) and adjust the weights of the samples (e.g., assigning higher weights to areas with missing data) to balance the distribution of the treatment and control groups. Outcome regression models can correct for biases caused by measurement errors or confounding factors (such as abnormal soil moisture data due to sensor malfunction), ensuring that feature values ​​are consistent with the actual physical processes. The synergistic mechanism of the dual robustness algorithms effectively eliminates bias; even if one model (propensity score or outcome regression) has an error, the other model can still guarantee the accuracy of the overall estimate, reducing the risk of single model failure. After processing by the dual robustness algorithms, the features in the matrix have been freed from bias and can truly reflect the causal relationships between factors (e.g., the direct impact of "rainfall → runoff"), rather than simply statistical correlation.

[0051] Temporal alignment can unify monitoring data from different time points (such as rainfall data every 10 minutes or reservoir water level data every hour) to the same time interval (such as 10 minutes), ensuring the consistency of time series. Spatial alignment can map spatial data from different sources (such as grid data from weather radar or point data monitored by BeiDou) to a unified spatial coordinate system (such as WGS84), and generate regular grids through interpolation and other methods, which is convenient for subsequent causal graph modeling (such as Thiessen polygon mesh generation).

[0052] Compared to traditional principal component analysis (PCA) which only reduces dimensionality through statistical correlation, this step uses a dual robust algorithm to achieve feature fusion through physical mechanism constraints (such as propensity scores corresponding to disaster occurrence probabilities and result regressions corresponding to hydrological processes). This reduces the soil moisture content prediction error from 18% to 7%, significantly improving data quality and laying the foundation for subsequent dynamic causal graph modeling.

[0053] Step 2: Dynamic Cause-Effect Graph Modeling and Weight Update The operation objects of this step are the feature matrix F(t) output from step 1 and the historical disaster dataset D={X,T,Y}.

[0054] The processing steps include: (1) Determining causal orientation based on additive noise model (ANM): like Y = f ( X )+ e and e ⊥ X Then establish X → Y The causal edges (HSIC test p-value < 0.01), specifically, the logic for determining the causal direction based on ANM is as follows: If variable Y Can be represented as a variable X The function and independent noise e The sum (i.e. Y = f ( X )+ e ), and noise e and X Statistical independence (denoted as e ⊥ X If ), then it is determined that there exists from X arrive Y causal relationship (i.e., establishing) X → Y (causal edges). Independence is verified using the HSIC test (Hilbert-Schmidt independence criterion test). p When the value is <0.01, it is considered that... e andX The independent hypothesis holds true, thus confirming the causal direction.

[0055] (2) Quantifying causal chains using probability tree reasoning:

[0056] An early warning is activated when the probability of transmission along a critical path (such as rainfall → runoff) is ≥0.8.

[0057] The probabilistic tree reasoning of this invention constructs a tree structure to represent the transmission path of causal relationships (e.g., "rainfall → soil saturation → landslide → river blockage → water level surge") and calculates the conditional probability between nodes on each path. Its function is to decompose complex causal relationships of flash floods into quantifiable chain structures, such as the direct causal chain of "rainfall → runoff," or an indirect causal chain containing multiple intermediate nodes (e.g., rainfall → increased soil moisture content → decreased permeability → surge in surface runoff); and to form the transmission probability of causal chains by statistically analyzing the transfer probabilities between nodes using historical disaster data (e.g., "the probability that heavy rainfall will cause runoff to exceed the warning value is 75%").

[0058] The critical path is the causal chain that plays a decisive role in the occurrence of flash floods, such as "heavy rainfall → soil saturation → landslide blocking river channels → water level surge." The interruption or regulation of this type of path can significantly affect the disaster outcome. When the total transmission probability of a causal path exceeds 80%, it indicates that the path is highly likely to trigger a disaster under current conditions, and the system will automatically activate an early warning.

[0059] This invention transforms the traditional "qualitative experience judgment" into a "quantitative data-driven" early warning mechanism by probabilistically quantifying the causal chain, thereby avoiding false alarms or missed alarms caused by a single indicator threshold (such as looking only at rainfall) and improving the scientificity and reliability of the early warning.

[0060] (3) Dynamically updating causal weights: dynamic parameter adjustment and data-driven mechanism

[0061] This indicates the change in water level per unit time. c The proportion of the contribution of water level changes to the weight is determined.

[0062] The weight of each edge in the cause-effect graph (e.g.) w ij ( t Weights represent the strength or degree of influence of a causal relationship, such as the "weight of the impact of rainfall on runoff" or the "weight of the impact of landslides on river blockage." The purpose of dynamically updating weights is to allow the model to adapt to environmental changes in real time (such as topographic changes and the status of engineering facilities), avoiding model lag caused by static weights.

[0063] Deformation rate Δ d ( t It can be calculated using BeiDou monitoring data (sampling frequency 1Hz); it can also be calculated using the differential interferometry technology of BeiDou satellites.

[0064] Water level change rate parameter c =0.63 (calibrated using historical dam failure data). c =0.63 is an empirical coefficient obtained by fitting historical dam break data, used to quantify the influence of "water level change rate" on causal weights.

[0065] The output is: a cause-effect graph with time-series weights. G ( t It contains ≥3 key causal paths (such as "heavy rainfall → soil saturation → landslide blocking river channel → water level surge").

[0066] Deformation rate is a precursor indicator of secondary disasters such as landslides and collapses. The higher the value, the more the causal weight of "topography → disaster" needs to be increased. For example, when the deformation rate in a certain area suddenly increases from 0.5 mm / h to 3 mm / h, the system will automatically increase the weight of "topography deformation → river blockage", indicating that the risk in the area has increased.

[0067] The dynamic weight update of this invention reduces the prediction error of flash flood time from ±25 minutes to ±8 minutes compared to a static Bayesian network.

[0068] The weight update frequency of this invention is synchronized with the BeiDou data sampling frequency (1Hz), that is, the causal graph weights are updated once per second to ensure that the model reflects the latest disaster situation in real time.

[0069] This invention provides direction for weight updates through probabilistic tree inference: by identifying critical paths with high transmission probabilities, it determines causal edges that require significant weight adjustments (e.g., "landslide → river blockage"). Dynamic weight updates enhance the accuracy of probabilistic inference: real-time weight adjustments allow the probability tree to more accurately reflect the strength of causal relationships in the current environment. For example, when surface deformation intensifies after an earthquake, weight increases can raise the transmission probability of the "earthquake → mountain loosening → landslide" path, avoiding underestimation of risk due to fixed weights. This "quantitative assessment + dynamic correction" mechanism enables the causal graph model to capture both the long-term statistical patterns of flash floods and adapt to real-time changes in the geographical environment, solving the core problems of traditional static models such as "strong lag and poor adaptability."

[0070] Step 3: Counterfactual scenario generation and multi-objective optimization This step operates on the current disaster situation. S ( t ), set of intervention measures A (such as the opening degree of the floodgate and the transfer route).

[0071] The processing steps include: (1) Generative Adversarial Models (CF-GANs): ,

[0072] The generator G uses a U-Net structure (hidden layer dimension = 256).

[0073] Discriminator D introduces a spectral normalization constraint (Lipschitz constant ≤ 1).

[0074] In CF-GANs, CF stands for Counterfactual, meaning the network is used to simulate potential disaster outcomes under different intervention measures (such as changing reservoir discharge or adjusting evacuation routes), belonging to the "counterfactual inference" technique in causal inference. Specifically, it simulates intervention scenarios that have not actually occurred through a generative model (generator G), helping to evaluate the effectiveness of different contingency plans and solving the problem that traditional methods "cannot quantify the causal impact of intervention measures".

[0075] U-Net is a classic neural network architecture originally used for image segmentation. Its key feature is a symmetrical encoder-decoder structure, fusing features from different layers via skip connections. In this invention, the input to the U-Net structure is the current disaster situation. S ( t ) and intervention measures A (e.g., floodgate opening degree, relocation route parameters), which may include spatiotemporal features (e.g., rainfall distribution, terrain data); the output is the generated counterfactual result. YCF Simulated data, such as water level changes and the distribution of affected areas, are used to evaluate the effectiveness of contingency plans. The hidden layer dimension is 256, meaning the number of neurons in the hidden layer of the network is 256, which determines the complexity of the model; this parameter balances computational efficiency and feature representation ability, enabling the network to capture the nonlinear causal relationship of the rainfall-runoff-disaster chain.

[0076] Spectral normalization is a regularization technique used to limit the spectral norm (maximum singular value) of the discriminator's weight matrix, thereby controlling its Lipschitz constant. In this invention, the discriminator D's task is to distinguish between real disaster data and simulated data output by the generator G. Through spectral normalization constraints, the discriminator is forced to evaluate data distribution differences in a smoother way, improving the quality of generated data (e.g., simulated disaster results are closer to real causal logic), and thus improving the accuracy of counterfactual inference. A Lipschitz constant ≤ 1 is a key condition for Wasserstein GAN (WGAN), ensuring the continuity of the discriminator's gradient Lipschitz, avoiding gradient explosion or vanishing, and making training more stable.

[0077] (2) Evaluation of the effectiveness of the causal distillation tree plan:

[0078] Split Criterion: Maximize the difference in treatment effects between groups (threshold) ITE≥0.15).

[0079] Causal distillation trees are used to quantify the causal effects of different interventions (such as reservoir scheduling plans and evacuation routes) on flash flood disasters. They can "distill" the causal relationships between key interventions and outcomes from complex causal graph models, and evaluate the effectiveness of different contingency plans.

[0080] Between-group differences in treatment effect: The treatment effect (ITE) refers to the degree of influence of a single intervention (such as opening a floodgate) on a specific outcome (such as the magnitude of water level drop). (The difference between the expected outcomes of the intervention group and the expected outcomes of the control group); maximizing between-group outcomes. ITE splitting criterion: When constructing tree nodes for causal distillation, the differences in treatment effects between the intervention and control groups are calculated based on features (such as rainfall and soil moisture), and then the nodes are selected. Features with an ITE ≥ 0.15 are split to ensure that the samples in each subtree have significant differences in causal effect; a feature is considered meaningful for the assessment of the plan only when the difference in the effect of the intervention on the outcome exceeds 0.15, thus avoiding the introduction of noisy features.

[0081] This invention visualizes the causal effects of different intervention paths through a tree structure, helping to quickly identify high-impact factors (such as the contribution of the "heavy rainfall + landslide" combination to the risk of dam failure), and providing a clear direction for contingency plan optimization.

[0082] (3) Pareto optimal solution selection: min[f1(casualties), f2(economic losses)] st Dam failure risk ≤ 0.1 An improved NSGA-II algorithm was used (crossover probability = 0.9, mutation probability = 0.05).

[0083] The output is: an optimal set of contingency plans containing 3-5 non-dominated solutions, with key indicators (such as expected transfer time and property loss range) marked for each plan.

[0084] The NSGA-II algorithm is a non-dominated sorting genetic algorithm that searches for the optimal solution set of multiple conflicting objectives by simulating the biological evolution process (selection, crossover, and mutation). The improvement in this invention lies in the fixed values ​​of the crossover and mutation probabilities. Crossover probability = 0.9: high probability of gene crossover operation, promoting feature fusion among individuals in the population and accelerating the generation of high-quality solutions; Mutation probability = 0.05: low probability of random mutation, maintaining population diversity and preventing the algorithm from prematurely converging to a local optimum.

[0085] The objective functions of Pareto optimal solution selection are minimizing casualties (f1) and minimizing economic losses (f2). These two core optimization objectives reflect the dual requirements of prioritizing life and protecting property. The constraint is that the risk of dam failure is ≤0.1, ensuring that the contingency plan is implemented within a controllable risk range.

[0086] The Pareto optimal solution is defined as follows: in multi-objective optimization, if no other solution can simultaneously outperform all objectives of the current solution, then the solution is called the Pareto optimal solution. This invention generates a set of contingency plans containing 3-5 non-dominated solutions, allowing decision-makers to weigh and choose based on actual scenarios (such as resource priorities and real-time disaster situations).

[0087] The advantage of this invention lies in its ability to achieve a counterfactual inference accuracy of 92%, which is 37% higher than traditional simulation; multi-objective optimization reduces cross-departmental resource conflicts by 89%.

[0088] Step 4: Multi-departmental collaborative execution and dynamic adjustment The objects of this step are the contingency plan set and real-time disaster feedback data generated in step 3.

[0089] The processing includes: (1) Two-layer reinforcement learning scheduling: Hydraulic layer: State space = {Reservoir storage W(t), Rainfall forecast R(t+)} The action space flood discharge Q is updated through the DQN network (ε-greedy strategy, ε=0.1).

[0090] Flood discharge optimization based on DQN employs the DQN algorithm (Deep Q-Network) from Deep Reinforcement Learning (DRL) to optimize the reservoir's water storage capacity. W (t ")" and "Future Rainfall Forecast" R ( t + t The state input is ")", and the output is the optimal flood discharge volume Q. The ε-greedy strategy (ε=0.1): During decision-making, the model selects the currently known optimal action (utilization) with a 90% probability and explores new actions randomly (exploration) with a 10% probability, avoiding the model getting trapped in local optima and adapting to real-time changes in water conditions. For example, when heavy rainfall is detected in the next 2 hours and the reservoir's water level is close to the warning level, DQN automatically calculates the optimal flood discharge volume using a strategy trained on historical data, thus avoiding the risk of upstream dam failure while reserving flood discharge space downstream.

[0091] Emergency Layer: Reward Function , m =0.7 (Congestion Weight) The emergency response layer of this invention employs a transfer optimization based on a reward function. This represents the total time required for the evacuation of people in each area. T ij For the first i area to number j The time at the resettlement site A ij (A 0-1 variable representing whether to choose this route) needs to be minimized. max( C ( t ) indicates the maximum level of congestion on the transportation network. m =0.7 indicates that congestion has a high weight, and priority is given to avoiding transfer delays caused by route congestion.

[0092] The goal of the emergency response layer is to balance "transfer speed" and "route accessibility" during emergency response. For example, it prioritizes routes with low congestion probability but longer distances to avoid overall transfer stalls due to local congestion.

[0093] This invention enables the linkage between the water conservancy layer and the emergency layer through reinforcement learning to coordinate "flood discharge scheduling" and "mass evacuation" - the water conservancy layer discharges floodwater in advance to lower the water level, buying time for the emergency layer to evacuate; the emergency layer adjusts the evacuation route according to the real-time water situation to avoid conflicts with downstream waterlogged areas caused by flood discharge and reduce resource conflicts.

[0094] (2) LBS fence trigger command: Tyson polygon grid division (side length ≤ 2km): The disaster-stricken area is divided into Tyson polygon grids with a side length of no more than 2 kilometers. Each grid corresponds to a geographical unit, realizing "grid management".

[0095] Compared to traditional administrative divisions, Tyson polygons are more closely aligned with actual geographical conditions (such as terrain and road network distribution), ensuring that the distance from the monitoring point to the center point within each grid is minimized, thus improving the spatial accuracy of command triggering.

[0096] Triggering conditions: or

[0097] II ( t >0.8 II crit This represents the "Comprehensive Disaster Impact Index" (a weighted sum of indicators such as water depth and flow velocity). When it exceeds 80% of the critical value, it indicates that danger is imminent in the area, triggering an early warning.

[0098] S soi ( t () represents soil moisture content. S sat The soil moisture content is saturated. When the moisture content exceeds 70%, the soil is prone to landslides or a surge in runoff, and emergency measures (such as blocking roads and organizing evacuations) must be initiated immediately.

[0099] By combining spatial grids with threshold monitoring, an upgrade from "global early warning" to "precise local triggering" is achieved. For example, if the soil moisture content in a certain grid reaches the threshold, but adjacent grids do not, a relocation order is only sent to the village within that grid, avoiding resource waste caused by over-response.

[0100] (3) Digital twin verification: Model correction is triggered when the error rate is >15%. Model calibration mechanism:

[0101] Y sim Simulation results, Y real According to actual monitoring data, when the error rate (such as root mean square error) exceeds 15%, it indicates that there is a significant deviation between the model prediction and the actual disaster situation, and correction needs to be triggered.

[0102] The output results are: cross-departmental instruction synchronization delay ≤ 2 minutes, and the success rate of contingency plan execution increased from 68% to 93%.

[0103] Traditional static emergency plans cannot adapt to sudden changes (such as unexpected rainfall or temporary bridge blockages). This invention's digital twin, through real-time data closure, achieves minute-level iteration of "monitoring-simulation-correction," ensuring the plan always matches the latest disaster situation. Through collaborative scheduling using two-layer reinforcement learning and precise LBS triggering, it breaks down data barriers between departments such as water resources, emergency response, and transportation, achieving efficient flow of instructions from generation to issuance, significantly improving upon the 15-minute delay of traditional methods. The dynamic correction mechanism ensures the plan continuously adapts to changes in the disaster situation during execution, avoiding execution failures due to initial model errors or unforeseen circumstances, significantly improving the reliability of emergency response.

[0104] Table 1

[0105] Any part of this invention not described in detail can be referred to in the prior art or in the art known to those skilled in the art. This embodiment does not limit such part and will not describe it in detail here.

[0106] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A method for generating dynamic emergency response plans for flash floods based on causal inference, characterized in that, The steps of the method include: S1. Obtain multi-source data and extract causal features from it; S2. Based on the extracted causal features and historical disaster datasets, a dynamic causal graph model is constructed and the weights are updated to obtain a causal graph with time-series weights. S3. Determine the current disaster status based on the cause-effect diagram information, and generate counterfactual contingency plans by combining the preset set of intervention measures to obtain the optimal contingency plan set; S4. Based on the optimal set of contingency plans, conduct multi-departmental collaborative execution.

2. The method for generating dynamic emergency plans for flash floods based on causal inference according to claim 1, characterized in that, The multi-source data includes meteorological radar data, BeiDou surface deformation monitoring data, reservoir scheduling logs, and traffic network status sensor data.

3. The method for generating dynamic emergency plans for flash floods based on causal inference according to claim 1, characterized in that, The causal feature extraction process in step S1 includes: Establish a spatiotemporal fusion matrix: calculate the terrain weight coefficient using the analytic hierarchy process and optimize the soil moisture sensitivity parameters using gradient descent. Data bias elimination: A robust bias elimination is achieved based on both a propensity score model and an outcome regression model; the propensity score model is implemented using an XGBoost classifier, and the outcome regression model is implemented using a random forest regression model. Output: The spatiotemporally aligned causal feature matrix is ​​obtained based on the determined causal relationships and then output.

4. The method for generating dynamic emergency response plans for flash floods based on causal inference according to claim 1, characterized in that, Step S2 includes: Determining the causal direction: Based on the additive noise model, if the variable Y Can be represented as a variable X The function and independent noise ε The sum of, and noise ε and X If statistical independence is established, then... X → Y Causal edge; Quantify causal chains: Use probability tree reasoning to determine whether the transmission probability of critical paths meets preset requirements, and activate an early warning when it does; Update causal weights: Calculate the deformation rate based on monitoring data, then calibrate the water level change rate parameter by combining it with historical dam break data, and dynamically update the weights based on the water level change rate parameter.

5. The method for generating dynamic emergency plans for flash floods based on causal inference according to claim 4, characterized in that, Determine noise ε and X The method for statistical independence is to verify it using the HSIC test. When the HSIC test shows that... p When the value is < 0.01, the two are considered to be independent.

6. The method for generating dynamic emergency plans for flash floods based on causal inference according to claim 1, characterized in that, Step S3 includes: Generative Adversarial Simulation: Adversarial simulation is performed using a generative adversarial network. The generator adopts a U-Net structure, and spectral normalization constraints are introduced into the discriminator. Evaluation of the effectiveness of the contingency plan: Causal distillation trees were used to evaluate the effectiveness of the contingency plan in order to maximize the difference in treatment effects between groups; Optimal solution selection: Pareto optimal solution selection is performed using the improved NSGA-II algorithm, with the objective of minimizing casualties and economic losses, and the constraint being the risk of dam failure.

7. The method for generating dynamic emergency plans for flash floods based on causal inference according to claim 1, characterized in that, Step S4 includes: Two-layer reinforcement learning scheduling: the water conservancy layer uses a DQN network to update the flood discharge volume, and the emergency response layer performs transfer optimization based on the reward function; LBS fence trigger command: Achieve precise response based on spatial grid by dividing the space into Thiessen polygonal grids; Digital twin verification: Dynamic model correction based on real-time data is performed through a model correction mechanism.

8. The method for generating dynamic emergency response plans for flash floods based on causal inference according to claim 7, characterized in that, The model calibration mechanism includes: determining the learning rate using gradient descent and adjusting model parameters using the Euclidean distance between the predicted and true values.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method for generating dynamic emergency plans for flash flood disasters based on causal inference as described in any one of claims 1-8.

10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the method for generating dynamic emergency plans for flash flood disasters based on causal inference as described in any one of claims 1-8.

Images