Monitoring and early warning method and system based on water conservancy flood prevention

Abstract

The invention relates to the technical field of water conservancy flood prevention monitoring, and discloses a water conservancy flood prevention monitoring and early warning method and system. The method comprises the following steps: acquiring multi-source hydrological monitoring data including real-time rainfall, riverway water level, soil humidity and weather forecast data, and constructing a comprehensive hydrological information basis; flood risk analysis is carried out based on the data, a flood risk prediction signal and a flood situation evolution signal are synchronously generated, and the flood occurrence probability and the development trend are clarified. And combining the flood risk prediction signal with the multi-source data to determine flood prevention regulation and control parameters and generate an optimized flood prevention signal. Flood prevention facility operation is adjusted according to the optimized flood prevention signal and the flood situation evolution signal, an early warning control signal is output, the two signals are processed at the same time to generate a comprehensive early warning index, and the severity of the flood situation is visually reflected. According to the method, hydrological data multi-dimensional fusion is realized, flood prevention monitoring and early warning are promoted to change from single data dependence to multi-source collaboration, and the flood situation response efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of water conservancy flood prevention monitoring, in particular to a water conservancy flood prevention monitoring and early warning method and system. BACKGROUND

[0002] The uneven spatiotemporal distribution of water resources and the frequent occurrence of extreme weather events make flood disasters an important factor threatening people's lives and property safety and restricting the sustainable development of social economy. As the key defense line against flood disasters, the scientificity and timeliness of water conservancy flood prevention monitoring and early warning are directly related to the effectiveness of flood prevention and disaster reduction. With climate warming and the acceleration of urbanization, flood disasters are characterized by suddenness, expansion of influence, and intensification of damage. The traditional flood prevention monitoring and early warning methods have gradually exposed many shortcomings and are difficult to meet the prevention and control needs under complex flood conditions.

[0003] Traditional flood prevention monitoring and early warning methods often rely on a single type of hydrological data, such as using only real-time rainfall or river water level as the core basis for judgment. The limitations of data sources lead to an incomplete understanding of flood conditions. Some areas still use a combination of manual monitoring and simple data statistics, which not only has low monitoring efficiency and lagging data updates, but also is easily affected by human error, affecting the accuracy of monitoring results. Even in some areas that have achieved automated monitoring, the data processing methods are mostly limited to simple threshold judgment, triggering an early warning only when the monitoring data reaches the preset threshold. This method cannot fully exploit the flood development rules hidden in the data and is difficult to make an early prediction of flood risks.

[0004] The application of meteorological forecast data in flood prevention work has obvious shortcomings. Most traditional methods fail to effectively integrate short-term and medium-term meteorological forecast information with real-time hydrological monitoring data, resulting in insufficient prediction accuracy of the time, magnitude, and influence range of floods. Soil moisture, as an important factor affecting surface runoff formation, is often neglected in traditional monitoring systems. The saturation state of soil directly determines the proportion of rainfall converted into surface runoff, and neglecting this data will lead to biased analysis of flood runoff mechanisms, affecting the pertinence of flood control measures.

[0005] In the flood response process, the traditional method lacks close connection between the early warning signal and the operation of flood control facilities. The early warning information is often presented in the form of simple text or numerical values, and cannot give clear facility control guidance combined with the evolution trend of flood conditions, which makes it difficult for flood control workers to make reasonable decisions quickly when facing complex flood conditions, and the role of flood control facilities cannot be fully played. In addition, the information output by the traditional early warning method is scattered, and there is no quantitative index that can comprehensively reflect the severity of the flood, making it difficult for flood control departments at all levels to quickly grasp the core of the flood, affecting the coordination efficiency of flood control resource allocation and emergency response. The existence of these problems leads to the situation that the traditional flood control monitoring and early warning method often fails to provide comprehensive and reliable technical support for flood control work when dealing with complex and variable flood disasters. SUMMARY

[0006] The purpose of the present application is to provide a water conservancy flood control monitoring and early warning method and system to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides a water conservancy flood control monitoring and early warning method, which comprises: obtaining multi-source hydrological monitoring data, the multi-source hydrological monitoring data including real-time rainfall data, river water level data, soil moisture data and weather forecast data; performing flood risk analysis based on the multi-source hydrological monitoring data to generate flood risk prediction signals and flood evolution signals; determining flood control regulation parameters according to the flood risk prediction signals and the multi-source hydrological monitoring data to generate optimized flood control signals; adjusting the operation of flood control facilities according to the optimized flood control signals and the flood evolution signals, and outputting early warning control signals; processing the flood risk prediction signals and the flood evolution signals to generate a comprehensive early warning index.

[0008] Preferably, the multi-source hydrological monitoring data includes: deploying an intelligent hydrological monitoring node network, the intelligent hydrological monitoring node network integrating rain gauges, water level sensors and soil moisture probes to collect the real-time rainfall data, the river water level data and the soil moisture data in real time; at the same time, the weather forecast data is obtained through weather radar scanning and geographic information system; the collected raw data is subjected to noise filtering and outlier rejection to form a standardized data stream; the standardized data stream is subjected to time and space alignment and feature extraction using a data fusion engine to generate the multi-source hydrological monitoring data, which is divided into a prediction data set, an optimization data set and a regulation data set.

[0009] Preferably, the flood risk analysis based on the multi-source hydrological monitoring data adopts a deep belief network model to process the prediction data set, the prediction data set including historical flood pattern data, real-time water flow velocity data and rainfall distribution map data; the historical flood pattern data is input into the deep belief network model for unsupervised pre-training to obtain feature representation; the real-time water flow velocity data and the rainfall distribution map data are combined to fine-tune model parameters through a back propagation algorithm, and the flood risk prediction signal is output; at the same time, the real-time water flow velocity data and the rainfall distribution map data are fused by using a Bayesian inference framework to calculate a posterior probability distribution and generate the flood evolution signal.

[0010] Preferably, constructing the deep belief network model includes a data normalization sub-step, a feature learning sub-step and a risk assessment sub-step; the data normalization sub-step performs minimum-maximum scaling on the historical flood pattern data to map the data to a unified interval; the feature learning sub-step extracts spatio-temporal features including rainfall cumulative features and river channel mutation features through a multi-layer restricted Boltzmann machine; and the risk assessment sub-step classifies the spatio-temporal features using a softmax classifier to output a flood risk level as the flood risk prediction signal.

[0011] Preferably, determining the flood control regulation parameter according to the flood risk prediction signal and the multi-source hydrological monitoring data includes: constructing a flood control decision knowledge graph that stores historical flood control cases and hydrological response relationships; inputting the optimization data set and the flood risk prediction signal into the flood control decision knowledge graph to generate an initial regulation scheme through graph neural network relationship reasoning; and iteratively optimizing the initial regulation scheme using a simulated annealing algorithm to adjust parameter weights and output the optimized flood control signal.

[0012] Preferably, constructing the flood control decision knowledge graph includes an entity recognition sub-step, a relationship extraction sub-step and a decision generation sub-step; the entity recognition sub-step extracts key entities including reservoir names and gate types from the historical flood control cases; the relationship extraction sub-step calculates the correlation strength between entities using a graph attention network to generate a flood control measure chain; and the decision generation sub-step dynamically searches for an optimal solution based on the flood control measure chain and the simulated annealing algorithm to update the optimized flood control signal.

[0013] Preferably, adjusting the flood control facility operation according to the optimized flood control signal and the flood evolution signal includes: Setting a flood safety boundary condition, the flood safety boundary condition is determined based on the regulation data set, the regulation data set includes real-time embankment stress data and flood propagation time data;Adaptive fuzzy control system is adopted to dynamically adjust the gate opening and pump station operation, the control quantity is calculated through membership function and rule base, and the early warning control signal is ensured to be maintained in the safe operation interval.

[0014] Preferably, the flood risk prediction signal and the flood evolution signal are processed to generate a comprehensive early warning index, comprising: constructing a decision support engine;The decision support engine applies grey correlation analysis method to process the flood risk prediction signal and the flood evolution signal, calculates the correlation degree sequence, and generates the comprehensive early warning index; The decision support engine contains multi-mode early warning state, corresponding to normal mode, alert mode and emergency mode respectively;When the risk value in the flood risk prediction signal exceeds the dynamic threshold value, switch to the alert mode, limit the output power of flood control facilities;When the flood evolution signal indicates that the flood peak is approaching, the emergency mode is activated, the standby flood control resources are started and the dispatch priority is adjusted.

[0015] Preferably, the optimized flood control signal includes target flood control area identification, resource allocation sequence, flow control curve and emergency response protocol.

[0016] Preferably, the application also includes a water conservancy flood control monitoring and early warning system based on the water conservancy flood control monitoring and early warning method, the system comprises a memory, a processor and a computer program stored in the memory and running on the processor, and the processor realizes the steps of the above-mentioned water conservancy flood control monitoring and early warning method when executing the computer program.

[0017] Compared with the prior art, the beneficial effects of the present application are: In the flood risk analysis link, instead of simply superimposing various data, the method generates flood risk prediction signal and flood evolution signal through cross validation and correlation analysis of multi-source data, and the synchronous output of the two signals realizes the double description of flood situation. The flood risk prediction signal focuses on the possibility of flood occurrence and the potential harm degree, and provides direction for early deployment of prevention and control measures;The flood evolution signal clearly presents the dynamic changes of the whole process from the formation to the development and subsidence of flood, so that the flood control workers can accurately grasp the rhythm of flood situation. This fine analysis of flood situation changes the traditional method which can only realize the rough judgment mode of "whether to warn", and realizes the accurate cognition of "when to occur, how to develop and how much harm" of flood situation, which creates conditions for the forward-looking deployment of flood control work.

[0018] The optimization of the generation process of the flood prevention signal fully embodies the scientificity and pertinence of the method. Instead of determining the regulation parameters according to fixed experience or a single threshold, the flood risk prediction signal is taken as the core orientation, and the real-time changes of multi-source hydrological monitoring data are combined to dynamically adjust the flood prevention regulation parameters. This process can fully consider the differences in hydrological characteristics of different regions, the runoff generation law under different rainfall intensities, and the influence of different soil moisture states on flood conditions, so that the generated optimized flood prevention signal is more in line with the actual flood condition demand. For example, in the region where the soil moisture has reached the saturation state, even if the rainfall does not reach the historical high value, the flood prevention facilities can be started in advance through parameter adjustment to avoid the rapid formation of surface runoff and cause floods; and in the dry soil region, the regulation threshold can be reasonably set according to the comprehensive judgment of data to avoid the waste of flood prevention resources.

[0019] The close connection between the early warning control signal and the operation of the flood prevention facilities effectively solves the problem of disconnection between early warning and execution in the traditional method. The method combines the optimized flood prevention signal with the flood condition evolution signal and directly converts it into specific flood prevention facility operation instructions, so that the early warning information is no longer an abstract numerical value or text, but an action instruction that can directly guide practice. Whether it is the opening and closing amplitude of the reservoir gate, the flow regulation of the flood drainage channel, or the key area of dike reinforcement, it can be clearly guided by the early warning control signal, greatly improving the execution efficiency and accuracy of the flood prevention measures. At the same time, the generation of the comprehensive early warning index realizes the high condensation of flood condition information, and converts the dispersed risk prediction signal, evolution signal, etc. into an intuitive quantitative index. The flood prevention departments at all levels can quickly judge the flood condition level through the index, clearly define their own responsibilities and work priorities, provide a unified information benchmark for cross-regional and cross-departmental flood prevention cooperation, and promote the efficient allocation of flood prevention resources and the orderly development of emergency response. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is the working principle diagram of the water conservancy flood prevention monitoring and early warning method described in the application; Figure 2 is the flowchart for obtaining multi-source hydrological monitoring data; Figure 3 is the flowchart for constructing a deep belief network model. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0022] Please refer to Figure 1The application provides a flood control monitoring and early warning method based on water conservancy. The method comprises the following steps: acquiring multi-source hydrological monitoring data, the multi-source hydrological monitoring data comprising real-time rainfall data, river water level data, soil moisture data and weather forecast data; performing flood risk analysis based on the multi-source hydrological monitoring data to generate a flood risk prediction signal and a flood situation evolution signal; determining a flood control regulation parameter according to the flood risk prediction signal and the multi-source hydrological monitoring data to generate an optimized flood control signal; adjusting the operation of flood control facilities according to the optimized flood control signal and the flood situation evolution signal, and outputting an early warning control signal; and processing the flood risk prediction signal and the flood situation evolution signal to generate a comprehensive early warning index. The multi-source hydrological monitoring data is acquired by relying on a distributed sensing network, the flood risk analysis is performed by using a machine learning model, the flood control regulation parameter is dynamically calculated by using an optimization algorithm, the operation of the flood control facilities is adaptively adjusted based on real-time signals, and the comprehensive early warning index is output by fusing multi-dimensional information by using a decision engine.

[0023] Embodiment 1: refer to Figure 2 The acquisition of multi-source hydrological monitoring data begins with a widely deployed intelligent hydrological monitoring node network. The intelligent hydrological monitoring node network covers key points of a target river basin in the physical space, such as riverbanks, reservoir dam bodies, and catchment areas. The intelligent hydrological monitoring node network integrates multiple types of sensor units, including high-precision tipping bucket rain gauges for collecting real-time rainfall data, pressure or ultrasonic water level sensors for continuously measuring river water level data, and time domain reflectometry soil moisture probes for acquiring soil volume water content data. These sensors work at a set sampling frequency, converting continuous physical hydrological phenomena into discrete digital signals to form raw monitoring data streams. The acquisition of weather forecast data relies on external system interfaces, connecting the radar scanning system and geographic information system of the meteorological department through a dedicated line to receive quantitative precipitation forecasts, satellite cloud images, and atmospheric circulation model data for a specific period in the future. Weather forecast data has a spatial and temporal grid feature, and its spatial resolution and time step need to match the ground monitoring data. During the process of transmitting raw data from sensors and external systems to the data aggregation center, various noise and interference will inevitably be introduced, including electronic device thermal noise, signal attenuation of transmission links, and occasional effects of environmental factors. A multi-level filtering and checking mechanism is designed for data preprocessing to clean the raw data. The noise filtering algorithm uses a digital filter to suppress high-frequency random fluctuations, and the outlier rejection logic identifies and removes outliers that are obviously outside the historical normal range based on statistical principles. The cleaned data enters the standardization processing procedure, which uniformly converts different dimensions and magnitudes of physical quantities into a dimensionless numerical interval, eliminating the bias caused by sensor differences and unit inconsistencies, and generating standardized data streams with consistent formats.

[0024] The data fusion engine is the core computing module that processes the standardized data stream. It has two major functions: spatio-temporal alignment and feature extraction. The spatio-temporal alignment function addresses the inconsistency of multi-source data in time stamp and geographic coordinate. The time alignment resamples the non-uniformly sampled data sequence to a uniform time grid through interpolation algorithm, while the space alignment fuses the point-like monitoring data and the surface-like prediction data into a unified geographic grid framework using the spatial interpolation technology of geographic information system. The feature extraction function mines meaningful patterns from the aligned data cube. The feature extraction operation includes calculating the rainfall accumulation in the sliding time window, identifying the rising inflection point of the water level sequence, and analyzing the spatial gradient distribution of soil moisture. The data fusion engine finally outputs the structured multi-source hydrological monitoring data, which is divided into three subsets with clear purposes according to the preset rules. The prediction data set mainly contains historical flood pattern data, real-time water flow velocity data, and high-resolution rainfall distribution map data. These data focus on reflecting the dynamic process and evolution law of the hydrological system. The optimization data set focuses on the current system state variables for real-time optimization decision. The regulation data set contains instantaneous parameters that directly affect the operation of the facility, such as the readings of the dike stress sensor and the flood propagation time calculated by the hydrological model.

[0025] The flood risk analysis module takes a prediction dataset as input, and its core is a deep belief network model. The deep belief network model is a kind of probabilistic generative model containing multiple hidden layers, and its training process is divided into two stages. The first stage is unsupervised pre-training, and the pre-training stage uses a large amount of historical flood pattern data, which records the hydro-meteorological conditions of the entire sequence from the beginning of the rainfall to the formation of the flood peak in past flood events. The pre-training process learns the high-order feature representation of the historical flood pattern data layer by layer through a multi-layer restricted Boltzmann machine. Each layer of the restricted Boltzmann machine attempts to capture the complex correlations in the input data with its hidden nodes. After multi-layer stacking, the model can extract abstract feature representations such as "continuous heavy rainfall superimposed on saturated soil" that lead to high-risk floods. The feature representation constitutes the deep understanding of the model of the causes of floods. The second stage is supervised fine-tuning, and the fine-tuning stage combines the pre-trained deep belief network model with the latest real-time observation data. Real-time flow velocity data is obtained from river flow meters, and rainfall distribution map data comes from meteorological radar inversion. These real-time data are input into the model. Through the error backpropagation algorithm, the actual occurred flood risk level is taken as the label, and all parameters of the deep belief network model are finely adjusted. The backpropagation algorithm calculates the difference between the model prediction output and the true label, and propagates this error signal from the output layer to the input layer layer by layer, according to which the connection weights and bias terms of each layer are updated. The fine-tuning process enables the deep belief network model not only to have the universal laws learned from history, but also to be adaptive to the current hydrological situation. Finally, a quantitative flood risk prediction signal is produced from the softmax classifier of the output layer, which is usually in the form of discrete risk levels or continuous probability values.

[0026] The generation of the flood evolution signal is parallel to the generation of the flood risk prediction signal. The generation of the flood evolution signal adopts a Bayesian inference framework. The Bayesian inference framework takes the real-time water flow velocity data and the rainfall distribution map data as new evidence and takes the flood evolution prior knowledge simulated by the hydrological and hydraulic model as the prior probability distribution. The Bayesian inference combines the prior probability with the likelihood function of the current evidence by the Bayes theorem to calculate the posterior probability distribution. The posterior probability distribution describes the possibility of the flood reaching a specific intensity at different times and different locations in the future under the given current observation condition. This constantly updated posterior probability distribution is materialized as the flood evolution signal, which can dynamically depict the advancing path of the flood wave, the predicted arrival time, and the uncertainty range of the peak intensity in the form of a spatiotemporal probability field. The deep belief network model and the Bayesian inference framework work together, one of which mines deep risk patterns from data and the other of which quantifies the uncertainty of the evolution process. The flood risk prediction signal and the flood evolution signal output by the two provide comprehensive and quantitative analysis basis for subsequent flood control regulation decisions. The deployment density and location optimization of the intelligent hydrological monitoring node network depend on the hydrogeological survey results, the algorithm parameters of the data fusion engine need to be calibrated through historical data, the structure design of the deep belief network model needs to strike a balance between model complexity and computational efficiency, and the selection of the prior distribution of the Bayesian inference framework needs to draw on domain knowledge.

[0027] Example 2: see Figure 3, the construction of DBN model is a structured process, which includes three sequentially executed sub-steps, namely data normalization sub-step, feature learning sub-step and risk assessment sub-step. The data normalization sub-step is the first step of the model processing input data, which preprocesses the historical flood pattern data. The historical flood pattern data usually comes from the long-term accumulated flood event records in the hydrological database. These original data have significant differences in numerical range, dimension and distribution. The data normalization sub-step adopts the min-max scaling algorithm, which linearly transforms the data of each feature dimension to a unified numerical interval, such as the range of zero to one. This transformation process is based on the minimum and maximum values of each feature dimension, and the original data points are converted to a new scale through mathematical mapping relationship. The data normalization sub-step makes the influence of different features on the model at a comparable level, avoiding the dominance of some features with large numerical range in model training, and also improves the numerical stability and convergence speed of the DBN model in the training process. The historical flood pattern data processed by the data normalization sub-step forms a standardized input matrix, preparing for the feature learning stage. The feature learning sub-step is the core of the DBN model, which is realized by constructing a multi-layer RBM stack structure. Each layer of RBM is a two-layer neural network, which contains a visible layer and a hidden layer. The visible layer is used to receive input data, and the hidden layer is used to learn the abstract representation of data. The training process of the feature learning sub-step is performed layer by layer greedily. The visible layer of the bottom layer RBM receives the normalized historical flood pattern data, and learns the activation values of the first layer hidden nodes through the contrastive divergence algorithm. These activation values constitute the first-order feature representation of the data, such as capturing the short-term rainfall intensity pattern. The hidden layer output of the first layer RBM is used as the input of the second layer RBM, and the second layer RBM continues to learn higher-level features, such as combining the temporal and spatial accumulation characteristics of rainfall from the first-order intensity pattern. In this way, the stack of multi-layer RBM enables the model to extract hierarchical feature representation from the original data. The high-level features are more abstract and can represent complex joint patterns, such as identifying the river mutation feature in the high-risk scenario of "continuous heavy rainfall in the upstream combined with high water level in the downstream river". The output of the feature learning sub-step is the final feature vector obtained after multi-layer nonlinear transformation, which condenses the most discriminative information in the input data.

[0028] The risk assessment sub-step is responsible for transforming the learned features into specific risk predictions. This sub-step utilizes a softmax classifier as its output layer. The softmax classifier receives the final feature vector generated by the feature learning sub-step as input. The softmax classifier itself is a multi-class logistic regression model. It calculates the probability of the input features belonging to each flood risk level through a linear transformation followed by a softmax activation function. The softmax activation function maps the scores after the linear transformation to a probability distribution between zero and one, with the sum of the probabilities of all risk levels being one. The class with the highest probability is the flood risk level predicted by the model, and this level signal is output as the flood risk prediction signal. The training objective of the deep belief network model is to minimize the cross-entropy loss between the predicted risk level and the true risk level label. Backpropagation and gradient descent optimization algorithms are used to adjust all model parameters from the restricted Boltzmann machine to the softmax classifier.

[0029] The method for determining flood control control parameters is based on a flood control decision-making knowledge graph, which is a knowledge base organized and stored in a graph structure. The flood control decision-making knowledge graph stores two types of core knowledge elements: historical flood control cases, which record the various flood control measures taken and their effects in specific flood events in the past; and hydrological response relationships, which describe the causal relationship between different water conservancy project control actions and changes in downstream hydrological elements. Nodes in the flood control decision-making knowledge graph represent entities, such as specific reservoirs, sluice gates, river sections, and monitoring stations, while edges represent relationships between entities, such as "Reservoir A affects the water level of downstream river section B through its spillway gate." The optimization dataset and flood risk prediction signals are jointly input into the flood control decision-making knowledge graph system. The optimization dataset contains the latest hydrological monitoring status information at the current moment. Graph neural networks are the core tool for querying and reasoning within the flood control decision-making knowledge graph. Graph neural networks can process the topological information of the graph structure and learn representations of nodes. The optimized dataset and flood risk prediction signals are mapped to the attribute features of specific nodes in a flood control decision-making knowledge graph. A graph neural network iteratively propagates information between nodes in the graph through a message-passing mechanism. At each layer, each node aggregates the feature information of its neighbors and updates its own representation by combining these features. After computation by multiple layers of the graph neural network, each node ultimately obtains an embedding vector that incorporates global graph context information. Based on these rich node representations, the system can perform relational reasoning, such as inferring which combination of control measures should be taken under the current high flood risk prediction signal and the water situation described by the optimized dataset. The result of relational reasoning is the generation of an initial control plan, which may include recommendations for flood discharge from multiple reservoirs, recommended gate opening levels, and recommended start-up and shutdown strategies for pumping stations.

[0030] Simulated annealing (SAM) is introduced to refine initial flood control schemes. It is a global optimization algorithm inspired by solid-state annealing. SAM uses the initial control scheme as a starting point and defines an objective function to evaluate its effectiveness. This objective function may comprehensively consider multiple goals, such as reducing flood risk, ensuring engineering safety, and minimizing economic losses. SAM searches for better solutions by performing random walks in the solution space. In each iteration, the algorithm generates a neighboring scheme based on the current scheme and calculates its objective function value. If the new scheme is better, it is accepted; if it is worse, it is accepted with a decreasing probability over time. This probability helps the algorithm escape local optima. Through this mechanism, SAM iteratively optimizes the complex flood control decision space, continuously adjusting the weights of various parameters in the initial control scheme, such as adjusting the priority of flood discharge from different reservoirs and optimizing the timing of gate openings. After multiple iterations and cooling processes, SAM finally outputs an optimized flood control scheme, which is encoded as an optimized flood control signal. Optimizing flood control signals includes specific and executable control instructions, thereby achieving precise control over water conservancy facilities. The feature learning capability of the deep belief network model, combined with the symbolic knowledge of the flood control decision knowledge graph, the reasoning capability of the graph neural network, and the optimization search capability of the simulated annealing algorithm, together constitute a complete technical chain from risk perception to decision generation.

[0031] Example 3: The entity recognition sub-step is the starting point for constructing a flood control decision-making knowledge graph. Its task is to automatically identify entities with specific meanings from unstructured historical flood control case text data. This historical flood control case text data originates from archived reports of flood control command departments, reservoir operation logs, and related technical summary documents. This text data records the environmental background, specific measures taken, and actual effects of each flood response process in natural language. The entity recognition sub-step employs named entity recognition technology based on a pre-trained language model. This model, trained on a large amount of labeled text data, can identify words or phrases representing specific categories in the text. In the field of flood control, the entity categories of interest mainly include reservoir name entities and gate type entities. Reservoir name entities are used to uniquely identify a specific water conservancy project facility, such as "Three Gorges Reservoir" or "Danjiangkou Reservoir"; gate type entities are used to classify and describe the structural form and functional characteristics of gates, such as "arc gate," "flat gate," or "flood discharge gate." The entity recognition sub-step processing flow includes text segmentation, part-of-speech tagging, entity boundary detection, and entity type classification. The final output is a structured list of entities extracted from the text, with each entity carrying its location information in the text and an assigned type label.

[0032] The relation extraction sub-step follows the entity recognition sub-step, and its goal is to establish semantic relationships between the entities identified in the entity recognition sub-step. This sub-step uses a graph attention network (GNN) deep learning model to accomplish this task. GNNs are neural networks specifically designed for processing graph-structured data, assigning different importance weights to nodes and edges in the graph. In the relation extraction sub-step, the entity set obtained from the entity recognition sub-step, along with the grammatical structure within sentences in historical flood control case texts, is initially constructed into a heterogeneous graph. The nodes in the graph are the identified entities, and the edges initially represent the co-occurrence relationships of entities appearing together in the same sentence. The GNN model is applied to this initially constructed graph, and it calculates the association strength between each node and its neighbors using a multi-head self-attention mechanism. For each pair of entity nodes with a potential relationship, the GNN learns an attention coefficient, the magnitude of which reflects the confidence that a specific semantic relationship exists between the two entities. For example, the GNN needs to determine whether the relationship between "Reservoir A" and "Gate B" is one of "equipped" or "controlled". The output of the relation extraction sub-step is a set of relations with explicit semantic labels between entities. These relation chains are interconnected, forming a flood control measure chain with practical business significance. The decision generation sub-step is the stage where the flood control decision knowledge graph generates the final decision output. Based on the flood control measure chain generated by the relation extraction sub-step, the decision generation sub-step uses a simulated annealing algorithm for dynamic search. The flood control measure chain provides a decision space, describing feasible sequences of control actions and their potential impacts. The decision generation sub-step maps real-time hydrological status information, such as optimization datasets and flood risk prediction signals, to the state attributes of corresponding nodes in the flood control decision knowledge graph. The core of the decision generation sub-step is defining an evaluation function to quantify the merits of any candidate decision scheme. The evaluation function needs to consider multiple factors, and its goal is to find an optimal combination of flood control control parameters. The simulated annealing algorithm is used for efficient search in this complex high-dimensional decision space. The simulated annealing algorithm starts from an initial solution, which can be generated based on routine operational suggestions in the flood control measure chain. During the algorithm's iteration process, simulated annealing randomly perturbs the current solution, generating a new candidate solution, and calculates the evaluation function value corresponding to the new solution. Simulated annealing decides whether to accept the new solution based on an acceptance probability that gradually decreases over time. Even if the new solution is worse than the current solution, it still has a certain probability of being accepted, which helps the algorithm escape local optima. The core of simulated annealing lies in the fact that its probability of accepting a new solution is determined by the following formula:

[0033] in: This represents the probability of accepting a worse new solution. The difference between the evaluation function value of the new solution and the evaluation function value of the original solution (new solution value - original solution value) is represented by k, which is an analogous parameter of Boltzmann constant in optimization algorithms, used to adjust the sensitivity of probability to energy difference. T is the temperature parameter of the current iteration step of the simulated annealing algorithm, which gradually decreases from a high initial value as the iteration progresses. Both sides of the formula are dimensionless probability values, maintaining consistent dimensions. The decision generation sub-step dynamically searches and updates the optimized flood control signals through continuous iteration of the simulated annealing algorithm, enabling the output decision scheme to adapt to the current complex and changing flood situation. The construction process of the flood control decision knowledge graph transforms unstructured textual experience into a structured knowledge network, and then, through the combination of graph neural networks and optimization algorithms, realizes the transformation from historical experience data to real-time scientific decision-making. The construction process of the flood control decision knowledge graph embodies an advanced decision-making paradigm that combines knowledge-driven and data-driven approaches. The model for the entity recognition sub-step needs to be fine-tuned to suit the technical terms used in flood control. The graph attention network structure design for the relation extraction sub-step affects the accuracy of relation extraction. The cooling schedule parameters of the simulated annealing algorithm in the decision generation sub-step need to be carefully set.

[0034] Example 4: Adjustments to flood control facility operations are based on flood control safety boundary conditions. These conditions are a set of limit parameters calculated using hydrodynamic and structural safety models. They define the permissible operational range of flood control facilities such as gates and pumping stations under extreme hydrological conditions, preventing engineering instability or secondary disasters due to improper operation. The determination of these boundary conditions relies on a control dataset, which includes real-time levee stress data and flood propagation time data. Real-time levee stress data is collected using strain gauges embedded within the levee, reflecting the mechanical response of the levee structure under the combined effects of hydrostatic pressure, seepage pressure, and wave scouring. Flood propagation time data is calculated using hydraulic formulas based on river topography and real-time flow velocity data, predicting the time required for flood waves to propagate from upstream key sections to important protected targets. The flood control safety boundary conditions compare these real-time data with the facility's design safety parameters, dynamically setting the upper and lower safety limits for facility operation, such as the maximum permissible water level, maximum gate opening, and maximum pumping station start-stop frequency.

[0035] The adaptive fuzzy control system is the core control unit for adjusting the operation of flood control facilities. It can handle the uncertainties and nonlinearities in flood control safety boundary conditions and real-time flood situation evolution signals. The design of the adaptive fuzzy control system comprises two core components: membership functions and a rule base. Membership functions are used to transform precise input variables into fuzzy linguistic values. Input variables include the deviation between the real-time river water level and the safe water level in the flood control safety boundary conditions, and the deviation between the flood propagation time and the preset emergency response time. Taking water level deviation as an example, five fuzzy sets can be defined: "negative large," "negative small," "zero," "positive small," and "positive large." Each fuzzy set is described by a membership function curve, such as a triangular function or a trapezoidal function. The rule base stores fuzzy control rules based on expert experience. These rules adopt an "if-then" form, such as "if the water level deviation is positive large and the flood propagation time deviation is negative small, then the gate opening adjustment is positive large." The workflow of an adaptive fuzzy control system is as follows: First, the precise input quantities are fuzzified using membership functions to obtain the membership degree of each input variable to each fuzzy linguistic value. Then, fuzzy inference is performed based on a rule base, where each rule generates a conclusion, which is also a fuzzy set. Finally, using defuzzification methods, such as the centroid method, the fuzzy conclusions derived from multiple rules are merged into a precise control output. This precise control output is the specific operational instruction, such as the adjustment amount of the gate opening or the setting value of the pump station speed. By periodically executing a "sampling-calculation-output" cycle, the adaptive fuzzy control system dynamically adjusts the gate opening and pump station operating status, ensuring that the facility operations driven by the early warning control signal are always maintained within the safe operating range defined by the flood control safety boundary conditions.

[0036] Referring to Table 1, the optimized flood control signal is the final output generated after optimization using the flood control decision-making knowledge graph and simulated annealing algorithm. The optimized flood control signal is a structured data object designed to precisely guide the spatial deployment, temporal scheduling, and operational details of flood control resources. The optimized flood control signal comprises four distinct components: target flood control area identifier, resource allocation sequence, flow control curve, and emergency response protocol. The target flood control area identifier precisely specifies the geographical area requiring flood control regulation. It typically employs standard coding systems from geographic information systems, such as watershed codes, administrative division codes, or custom flood protection zone numbers, ensuring unambiguous instructions. The resource allocation sequence details the scheduling order and allocation ratio of flood control materials and human resources. It is formulated based on the risk levels of each region assessed by flood risk prediction signals and flood situation evolution signals, following a risk-first principle. The flow control curve is the core control instruction in the optimized flood control signal. It describes, in time series form, the ideal flow process line expected to be achieved at key control sections of the target flood control area. The flow control curve is calculated by a hydraulic model based on upstream inflow, confluence within the affected area, and downstream capacity. Its purpose is to smooth flood peaks and stagger flood discharge. The horizontal axis represents time, and the vertical axis represents flow rate. The emergency response protocol specifies standardized operating procedures to be triggered when the flood situation reaches a certain stage. The protocol includes provisions for personnel evacuation routes, communication support plans, and backup power activation procedures.

[0037] Table 1: Key Parameters of Early Warning Control Signals

[0038] The optimized flood control signals are transmitted to the adaptive fuzzy control system and the terminal equipment of the flood control command department through a standardized data interface. Target flood control area identification ensures the targeted nature of control actions, resource allocation sequences optimize the utilization efficiency of limited resources, flow control curves provide quantifiable targets for the precise control of facilities such as gates, and emergency response protocols guarantee the orderliness and standardization of actions in emergency situations. The adjustment process of flood control facility operations and the definition of optimized flood control signals work together to form a closed-loop control link from macro-level decision-making to micro-level execution. The adaptive fuzzy control system is responsible for transforming the abstract targets in the optimized flood control signals into specific mechanical actions, while flood control safety boundary conditions provide a safety guarantee framework for the entire operation process. The monitoring frequency of real-time levee stress data matches the update cycle of the control system; the shape parameters of the membership function need to be tuned according to the control characteristics of specific facilities; the completeness of the rule base directly affects the performance of the control system; and the data structure design of the optimized flood control signals considers communication bandwidth and parsing efficiency.

[0039] Example 5: The decision support engine is the core computational module for generating the comprehensive early warning index. It receives flood risk prediction signals and flood situation evolution signals from the upstream analysis process as input data. The flood risk prediction signal is a quantitative indicator, possibly represented as a continuous value between zero and one; a higher value indicates a greater probability of flood occurrence or a greater expected degree of harm. The flood situation evolution signal contains richer spatiotemporal information, such as flood inundation depth forecasts expressed in grid form and estimated time series of flood peak arrivals at key cross-sections. The decision support engine needs to fuse these two signals of different dimensions and scales into a single, easily understood, and operable comprehensive index—the comprehensive early warning index. Grey relational analysis is the core algorithm used for signal fusion within the decision support engine. Grey relational analysis is suitable for handling small-sample systems with incomplete or unclear information. The decision support engine constructs a comparison sequence to be evaluated from the current flood risk prediction signal and the flood situation evolution signal, while defining an idealized reference sequence representing the most extreme and dangerous flood situation. The calculation steps of grey relational analysis involve comparing the geometric similarity between the comparison sequence and the reference sequence at each corresponding point. The closer the shapes are, the higher the correlation between the current situation and the extreme danger situation. The correlation calculation involves the absolute difference, minimum difference, and maximum difference of each data point, ultimately yielding a correlation coefficient sequence, where the correlation coefficient is a value between zero and one. The decision support engine performs a weighted average on this correlation coefficient sequence, with weights allocated based on the importance or reliability of the signals. The final weighted average is the comprehensive early warning index. As a scalar, the comprehensive early warning index comprehensively reflects the overall level of current flood risk. Its value range is typically defined between zero and one; a higher value indicates a more urgent flood situation and requires a higher level of response. The decision support engine includes multi-mode early warning states, a working paradigm that dynamically switches based on the comprehensive early warning index and specific signal characteristics. Multi-mode early warning states are mainly divided into three typical types: routine mode, alert mode, and emergency mode. Under normal conditions, the comprehensive early warning index is at a low level, the flood risk prediction signal shows that the risk value is far below the preset dynamic threshold, and the flood situation evolution signal does not indicate an imminent serious threat. Under normal conditions, the decision support engine's output focuses on routine monitoring and data recording, flood control facilities maintain basic operation, and resource allocation follows daily plans.

[0040] Dynamic thresholds are the key criterion for triggering mode switching. These thresholds are not fixed values; they are dynamically adjusted based on factors such as season, prior rainfall, and soil saturation. When the risk value in the flood risk prediction signal exceeds this dynamic threshold, the decision support engine automatically switches from normal mode to alert mode. Alert mode signifies that the system has identified a significant increase in flood risk, with the possibility of a more severe situation developing. In alert mode, the decision support engine outputs restrictive commands, limiting the output power of flood control facilities. For example, to prevent equipment from malfunctioning due to prolonged high load operation, the system may command pumping stations to operate at less than 80% of their maximum design capacity, or instruct gate operations to adopt a smoother opening rate to avoid sudden flow changes impacting downstream areas. Resource allocation in alert mode begins to favor high-risk areas, and on-duty personnel strengthen monitoring. Emergency mode represents the highest level of early warning. Activation of emergency mode is based on the judgment of flood situation evolution signals. When the flood situation evolution signal, calculated by the hydraulic model, indicates that the flood peak will reach the core protection area within a short period, such as two to three hours, the decision support engine immediately activates emergency mode. Emergency mode signifies an impending or already occurring flood disaster, requiring the most decisive response. In emergency mode, the decision support engine executes a series of pre-set emergency response protocols, including activating backup flood control resources and adjusting dispatch priorities. Activating backup flood control resources might mean immediately deploying mobile high-powered drainage pumps, and requisitioning strategic reserves of flood control bags and sand. Adjusting dispatch priorities is reflected in resource allocation, prioritizing the needs of key flood control projects and the smooth evacuation of personnel, while suspending all non-urgent routine maintenance activities. In emergency mode, flood control facilities operate to maximize their flood discharge or drainage capacity while ensuring safety.

[0041] The decision support engine's three multi-mode early warning states constitute a progressively stronger system. Taking a specific river basin as an example, in the early stages of rainfall, the comprehensive early warning index is low, and the system is in normal mode. As heavy rainfall continues, the calculated value of the flood risk prediction signal exceeds the dynamic threshold, and the system enters alert mode, with reservoirs beginning to release water to free up storage capacity. When the hydrological model predicts that the flood peak will reach a major downstream city in three hours, the flood situation evolution signal triggers emergency mode. The decision support engine commands all reservoir gates to open, activates the downstream flood diversion area, and issues the highest-level early warning information to the public. The decision support engine's grey relational analysis method effectively fuses multi-source heterogeneous signals, while the multi-mode early warning state mechanism ensures dynamic matching between flood control response measures and risk levels. The entire system's workflow embodies the core idea of ​​close coupling between situational awareness and hierarchical response in intelligent early warning. The coefficient resolution of the grey relational analysis method needs to be calibrated based on historical cases, the dynamic threshold adjustment algorithm needs to have adaptive learning capabilities, and the switching logic between multi-mode early warning states must be clear and unambiguous to avoid frequent oscillations near the critical point.

[0042] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A flood monitoring and early warning method for water conservancy, characterized in that, The method comprises: acquiring multi-source hydrological monitoring data, the multi-source hydrological monitoring data including real-time rainfall data, river water level data, soil moisture data and weather forecast data; performing flood risk analysis based on the multi-source hydrological monitoring data to generate a flood risk prediction signal and a flood evolution signal; determining flood control parameters according to the flood risk prediction signal and the multi-source hydrological monitoring data to generate an optimized flood control signal; adjusting flood control facility operation according to the optimized flood control signal and the flood evolution signal, and outputting an early warning control signal; processing the flood risk prediction signal and the flood evolution signal to generate a comprehensive early warning index.

2. The flood monitoring and early warning method of water conservancy according to claim 1, characterized in that, The acquisition of multi-source hydrological monitoring data comprises: deploying an intelligent hydrological monitoring node network, the intelligent hydrological monitoring node network integrating rain gauges, water level sensors and soil moisture probes to collect the real-time rainfall data, the river water level data and the soil moisture data in real time; at the same time, the weather forecast data is acquired through weather radar scanning and geographic information systems; the collected raw data is subjected to noise filtering and outlier rejection to form a standardized data stream; the standardized data stream is subjected to time-space alignment and feature extraction using a data fusion engine to generate the multi-source hydrological monitoring data, which is divided into a prediction data set, an optimization data set and a control data set.

3. The flood monitoring and early warning method of water conservancy according to claim 2, characterized in that, The flood risk analysis based on the multi-source hydrological monitoring data uses a deep belief network model to process the prediction data set, which includes historical flood pattern data, real-time water flow velocity data and rainfall distribution map data; the historical flood pattern data is input into the deep belief network model for unsupervised pre-training to obtain feature representation; combined with the real-time water flow velocity data and the rainfall distribution map data, the model parameters are fine-tuned through a back propagation algorithm to output the flood risk prediction signal; at the same time, the real-time water flow velocity data and the rainfall distribution map data are fused using a Bayesian inference framework to calculate the posterior probability distribution and generate the flood evolution signal.

4. The flood monitoring and early warning method of claim 3, wherein, The construction of the deep belief network model comprises a data normalization sub-step, a feature learning sub-step and a risk assessment sub-step; the data normalization sub-step performs minimum-maximum scaling on the historical flood pattern data to map the data to a unified interval; the feature learning sub-step extracts spatio-temporal features, including rainfall accumulation features and river mutation features, through a multi-layer restricted Boltzmann machine; the risk assessment sub-step classifies the spatio-temporal features using a softmax classifier to output a flood risk level as the flood risk prediction signal.

5. The flood monitoring and early warning method of claim 2, wherein, The determination of flood control parameters according to the flood risk prediction signal and the multi-source hydrological monitoring data comprises: The flood control decision knowledge graph is constructed, and the flood control decision knowledge graph stores historical flood control cases and hydrological response relationships; the optimized data set and the flood risk prediction signal are input into the flood control decision knowledge graph, relationship reasoning is performed through a graph neural network, and an initial regulation scheme is generated; the initial regulation scheme is iteratively optimized by using a simulated annealing algorithm, parameter weights are adjusted, and the optimized flood control signal is output.

6. The flood monitoring and early warning method of water conservancy according to claim 5, characterized in that, The flood control decision knowledge graph construction includes an entity recognition sub-step, a relationship extraction sub-step, and a decision generation sub-step; the entity recognition sub-step extracts key entities from the historical flood control cases, including reservoir names and gate types; the relationship extraction sub-step uses a graph attention network to calculate the correlation strength between entities and generates a flood control measure chain; and the decision generation sub-step dynamically searches for an optimal solution based on the flood control measure chain and the simulated annealing algorithm, and updates the optimized flood control signal.

7. The flood monitoring and early warning method of water conservancy according to claim 2, characterized in that, The adjustment of the flood control facility operation according to the optimized flood control signal and the flood situation evolution signal includes: A flood control safety boundary condition is set, which is determined based on the regulation data set, the regulation data set including real-time dike stress data and flood propagation time data; an adaptive fuzzy control system is used to dynamically adjust the gate opening and the pump station operation, and a membership function and a rule base are used to calculate the control amount, so that the early warning control signal is maintained within a safe operation interval.

8. The flood monitoring and early warning method of claim 1, wherein, The processing of the flood risk prediction signal and the flood situation evolution signal to generate a comprehensive early warning index includes: constructing a decision support engine; the decision support engine applies a grey correlation analysis method to process the flood risk prediction signal and the flood situation evolution signal, calculates a correlation degree sequence, and generates the comprehensive early warning index; The decision support engine includes multiple modes of early warning states, respectively corresponding to a normal mode, an alert mode, and an emergency mode; when the risk value in the flood risk prediction signal exceeds a dynamic threshold value, the alert mode is switched to, and the output power of the flood control facility is limited; when the flood situation evolution signal indicates that the flood peak is approaching, the emergency mode is activated, and standby flood control resources are started and the dispatch priority is adjusted.

9. The flood monitoring and early warning method of claim 1, wherein, The optimized flood control signal includes a target flood control area identifier, a resource allocation sequence, a flow control curve, and an emergency response protocol.

10. A flood monitoring and early warning system based on water conservancy, comprising a memory, a processor and a computer program stored in the memory and running on the processor, characterized in that, The processor, when executing the computer program, implements the steps of the flood control monitoring and early warning method based on the water conservancy flood control monitoring and early warning method according to any one of claims 1 to 9.

Images