Water conservancy sensor network flood disaster AI large model real-time management system

Abstract

According to the real-time management system for the flood disaster AI large model of the water conservancy sensor network, data access of flood sensors and exploration processing of flood events are based on an OGC SWE sensor network integration framework, and registration of the multi-source heterogeneous flood sensors and real-time access and storage of monitoring data are established; the diagnosis, preparation, response and recovery stages of flood events are explored under a unified standard, and flood early warning, hydrological prediction and flood event calculation services are issued in real time; a flood disaster AI large model, a flood event large model real-time service framework based on an SWE framework and a flood event large model real-time service system are established, and the problems of timely early warning, publishing and prediction of flood events can be solved; the BP neural network is adopted to predict the water level, meteorological factors such as air temperature, air pressure and atmospheric humidity are added on the basis of a rainfall and upstream water level prediction method, the prediction precision is improved, and high-timeliness, high-availability and high-fine-granularity flood information service is provided.

Description

Technical Field

[0001] The present application relates to a real-time management system for a large model of flood disasters, and in particular to a real-time management system for an AI large model of flood disasters in a water conservancy sensor network, belonging to the technical field of large model management of water conservancy flood disasters. Background Art

[0002] At present, how to dynamically display the flood status in real time and actively provide services has become an increasingly important topic. With the continuous development of sensor networks, the current disaster monitoring system has gradually changed from the original monitoring data service management mode to the event-driven automated real-time exploration and display mode. As a result, flood information systems have emerged in an endless stream, but the current flood information systems are mostly dedicated systems. Each flood information system is targeted at a specific research area, hydrological sensor type and service form. Since the general sensor network data standard is not used, each system is independent from each other at the data access layer, which affects the interaction and expansion between systems. In the exploration process, each system adopts coarse-grained flood exploration, and divides the flood system into two states: occurrence or not. Since the flood event object is not modeled in the flood exploration, each flood exploration system can only provide early warning services before the flood occurs, but cannot provide corresponding service content when and after the flood event occurs. Therefore, there are certain limitations in the expansion of flood information services. In addition, due to the lack of the "subscribe-publish" active service mechanism, the current flood exploration system has defects in regional expansion and timely flood warning for users.

[0003] The problems that need to be solved in the real-time management of water conservancy and flood disasters in the prior art and the key technical difficulties of this application include:

[0004] (1) How to dynamically display the flood status in real time and actively provide services is an urgent problem to be solved. With the continuous development of sensor networks, the current disaster monitoring system has gradually changed from the original monitoring data service management mode to the event-driven automated real-time exploration and display mode. However, the current flood information system is mainly based on dedicated systems. Each flood information system is targeted at a specific research area, hydrological sensor type and service form. Since the universal sensor network data standard is not used, each system is independent of each other from the data access layer, which affects the interaction and expansion between systems. In the exploration process, each system adopts coarse-grained flood exploration and divides the flood system into two states: occurrence or non-occurrence. Since the flood event object is not modeled in the flood exploration, each flood exploration system can only provide early warning services before the flood occurs, but cannot provide corresponding service content when and after the flood event occurs. Therefore, there are limitations in the expansion of flood information services. In addition, due to the lack of a "subscribe-publish" active service mechanism, the current flood exploration system has defects in regional expansion and timely flood warning for users.

[0005] (2) Due to the different service release forms, each flood information system operates independently, making it difficult to exchange flood information, resulting in the formation of information "islands". In addition, the current flood event information system simply divides flood events into two states: occurrence and non-occurrence, and cannot describe the stage process of flood occurrence, which leads to the inability of flood stages to evolve automatically. Therefore, the change of flood status is driven by the detection system or users, and the flood event itself cannot be actively changed. The reason is that the life cycle of flood events is not modeled, and the corresponding real-time flood detection model is not provided. In the face of disaster event management issues, the existing technology has not been able to provide an automatic detection model for flood events and a universal full-life cycle flood detection service method. It is urgent to comprehensively consider the hazards of flood disasters, the sharing and automatic detection of flood detection, and design and implement a real-time service system based on flood disaster information under a unified service framework, solve the data and service barriers of the current flood information system, provide more refined active services for flood events, and provide decision-making departments with more detailed flood disaster impact assessment and command decision support.

[0006] (3) Existing technologies divide flood detection into two categories: occurrence and non-occurrence. They fail to reflect the spatiotemporal state of floods from preparation to occurrence and final retreat, which makes them weak in active flood detection and early warning, and also has a negative impact on the management and sharing of flood events. Existing flood service methods are mostly based on SOA architecture and dedicated frameworks. Considering the real-time and sharing issues of flood events, a flood information active service system based on a general framework and a "subscribe-publish" model is needed. Existing technologies lack AI big models for flood disasters, lack a BP neural network water level prediction method that combines real-time flood detection and multiple meteorological monitoring data, lack a complete event stage that divides flood disasters into diagnosis, preparation, response and recovery, and cannot predict the changing trend of future flood events when detecting the current event status. There is a lack of a real-time service framework for big models of flood events, which cannot solve the process of real-time access, storage, release and processing of flood sensor data streams, and cannot publish flood status information in real time to actively provide flood information services. The lack of a real-time service system for large-scale flood event models makes it impossible to provide high-efficiency, high-availability and high-granularity flood information services. Summary of the invention

[0007] The real-time service system of the flood event big model based on the SWE framework proposed in this application, under the OGCSWE sensor network integration framework, uses the AI ​​big model of the flood event process to express the past, present and future changing characteristics of floods, explore the development process of flood events, and divide it into diagnosis, preparation, response and recovery stages. Compared with the traditional flood event exploration and release system, this system not only has good real-time performance, but also can solve problems such as data sharing, unified processing and exploration of multi-source data, and real-time active services. Based on the big model of flood event exploration, this system improves the fine-grainedness of flood event services, enriches the services at all stages of floods, provides users with flood warning, water level prediction and event disaster calculation services, provides real-time and active flood status analysis and decision-making support for relevant disaster prevention and control departments, prepares for reducing casualties and property losses in flood disasters, and provides decision-making departments with more detailed flood disaster impact assessment and command decision-making support.

[0008] In order to achieve the above technical effects, the technical solutions adopted in this application are as follows:

[0009] The real-time management system of AI big model of flood disaster in water conservancy sensor network builds a real-time service system of big model of flood event based on SWE framework. From data access of flood sensor to detection and processing of flood event, based on OGC SWE sensor network integration framework, it establishes real-time access and storage of multi-source heterogeneous flood sensor registration and monitoring data, and explores the diagnosis, preparation, response and recovery stages of flood event under unified standards, and releases flood warning, hydrological forecast and flood event calculation services in real time;

[0010] 1) Establish a large AI model for flood disasters: Based on hydrological and meteorological monitoring data, flood events are divided into four stages: diagnosis, preparation, response, and recovery. The stage status of flood events is calibrated, and the process stages of flood events are refined to provide a basis for early warning and high-granularity services for flood events. In addition, the establishment of a flood event process model includes two aspects: calculation and prediction. The calculation module explores flood information from historical and current hydrological monitoring data, and the prediction module predicts future flood change trends based on historical and current flood monitoring data.

[0011] 2) Establish a real-time service framework for a large model of flood events based on the SWE framework: Under the SWE standard, realize the automatic access of flood sensor monitoring data, automatic detection of flood events, and automatic triggering of flood stage services. Services are connected to each other based on data interfaces. The framework middle platform actively provides users with flood event detection results and related service content in the form of independent service operation. Flood warning, water level prediction and flood event calculation services are provided for the message content of the preparation, response and recovery stages of flood events. Users obtain flood status information actively pushed by the system through emails or network notifications. In terms of the scalability of multi-source data, two heterogeneous monitoring station data of meteorological stations and hydrological stations are connected. The system middle platform service ensures that the services in the framework work in coordination with the data interface, expands the low-level content of each service, and based on the active service mechanism of messages, ensures that the system can provide corresponding services according to the message status. Users expand different service contents according to the given interface specifications;

[0012] 3) Design of real-time service system for large-scale flood event model: Using flood data streaming access unit, based on Storm platform, real-time access to flood sensor monitoring data, and real-time release of sensor monitoring service, flood sensor monitoring data flows through SOSSESFeeder link service to sensor event server in real time, real-time detection of flood event status, based on real-time access to monitoring data, real-time detection of flood events, and real-time release of flood status through network notification service, triggering real-time active service after flood status change detection;

[0013] 4) Based on the BP neural network to predict water level, new meteorological factors such as temperature, air pressure, and atmospheric humidity are added to the prediction method based on precipitation and upstream water level, and the prediction accuracy should be improved.

[0014] Preferably, the AI ​​big model of flood disaster: the model is the rule for detecting flood events. It defines the factor characteristics of each stage of flood events based on the direct characterization factors of the current flood, including water level, flow and other information, and defines the information process model of flood events based on the characteristic pattern of water level changes. The AI ​​big model of flood disaster includes two parts. The first part is a characterization model for calculating whether the monitoring frequency of the water level monitoring value in a given interval is greater than or equal to the critical value within the past to the current time range. As a flood disaster calculation model, based on the water level as the direct characterization factor of flood, the flood event is divided into four stages: diagnosis, preparation, response and recovery, expressing the flood state within the time range from now to the past TW; the second part is a flood disaster prediction model for predicting water levels. Based on historical water levels and meteorological data, under the premise of factor normalization, the prediction model is trained based on the BP neural network to realize water level prediction. Based on the flood disaster calculation model in the first part, the future water level prediction function is provided when the flood disaster is in the preparation response and recovery stages. Based on the water level prediction, the future flood trend is analyzed;

[0015] Define W to represent water level data, U(N) to represent the critical water level interval of stage N, where N∈[A, B, C, D], A represents the diagnosis stage, B represents the preparation stage, C represents the response stage, and D represents the recovery stage. TW(N) represents the time window of stage N, and Fre(TW(N)) refers to the event frequency of the water level WEU(N) of stage N within the TW(N) time window. F(N) represents the critical value of the calculated frequency of stage N. The algorithm steps are as follows:

[0016] Step 1: Input water level monitoring data;

[0017] Step 2: When judging the water level W∈U(N), the value space P of N;

[0018] Step 3: Within the value range of P, calculate the frequency Fre(N) of the event type in the TW(N) time window and compare whether it is greater than F(N), where N∈P. If the calculated frequency is greater than the given value, record the object and further obtain the value space E of N.

[0019] Step 4: loop through each value in the value space E of N; if the value is A, output it directly; if the value is B, determine whether the state of the previous flood event is A, if so, output stage B, otherwise discard; if the value is C, determine whether the state of the previous flood event is B, if so, output stage C, otherwise discard; if the value is D, determine whether the state of the previous flood event is C, if so, output D, otherwise discard;

[0020] Step 5: According to the flood stage output in step 4, determine whether the flood event is in the preparation, response and recovery stages; if so, proceed to step 6 to provide flood disaster prediction services;

[0021] Step 6: Select the historical water level and meteorological data within the TW1 time range, normalize them to form training data, and train the BP neural network model; construct prediction data based on the current monitoring data within the TW2 time range, predict the water level, and obtain the prediction result R.

[0022] Preferably, the flood disaster calculation model is divided into diagnosis, preparation, response and recovery stage filtering conditions, and the exploration of each stage must meet two basic filtering conditions: hydrological monitoring data filtering and event stage dependency filtering;

[0023] Among them, each flood event stage needs to meet the basic hydrological monitoring data filtering conditions:

[0024] Condi=water level∈U&Fre(TW)≥F Formula 1

[0025] The frequency Fre of the water level monitoring data shown in Formula 1 appearing in a given water level interval U within a certain time TW is greater than or equal to the given frequency F;

[0026] Formula 2 is the dependent filtering condition of the flood event stage, which is the key to ensure the mutual dependence of each stage of the flood event. The definition of the flood event stage must strictly meet the sequential dependence relationship of A->B->C->D. The definition of the current stage CurrentPhase must be based on meeting the filtering conditions of the hydrological monitoring data and strictly depend on the previous exploration results;

[0027] CurrentPhase=Condi after astPhase, LastPhase∈[A, B, C] Equation 2

[0028] The filter condition A that the monitoring data in the diagnosis phase should meet is:

[0029] A=water level<W1&Fre(TW=T1)≥F1 Formula 3

[0030] The filter condition B that the monitoring data in the preparation phase should meet is:

[0031] BCondi=W1≤Water level<W2&Fre(TW=T2)≥F2

[0032] B=Bcondi after A Formula 4

[0033] The filtering condition C that the monitoring data in the response phase should meet is:

[0034] Ccondi=water level≥W2&Fre(TW=T3)≥F3

[0035] C=Ccondi after B Formula 5

[0036] The filtering condition D that the monitoring data in the recovery phase should meet is:

[0037] Dcondi=water level<W3&Fre(TW=T4)≥F4

[0038] D=Ccondi after C Formula 6

[0039] W1, W2 and W3 are water level values ​​at different stages, and W1<W2 and W3<W2 are satisfied; TW stands for TimeWindow, which refers to the time window. T1, T2, T3 and T4 are different critical values ​​of the time window, and T1≥T2≥T3≥T4 is satisfied; Fre(TW) refers to the number of times the qualified water level value appears in the time window TW, and F1, F2, F3 and F4 are different number values. The relative size restrictions of the water level value and the length of the time window are used to avoid confusion between different stages of the flood event and ensure the uniqueness of the stage of the flood event; in addition, the diagnosis, preparation, response and recovery stages are strictly in sequence. In addition to the need to judge whether the current basic conditions Bcondi, Ccondi and DCondi are met, the three conditions B, C and D also need to meet the change trend of the evolution of the A->B->C->D stages, that is, the preparation stage must be after the diagnosis stage, the response stage must be after the preparation stage, and the recovery stage must be after the response stage.

[0040] Preferably, the flood disaster prediction model: the calculation model characterizes the change characteristics of the hydrological monitoring data from the past to the present, and the prediction model performs hydrological prediction of flood events based on the flood event status information extracted after preliminary filtering by the calculation model, thereby improving the accuracy and timeliness of the re-judgment of the occurrence time and status of flood events;

[0041] The short-term water level prediction model based on BP neural network is composed of two parts: forward feedback and backward feedback. In the forward feedback process, the input vector X passes through the input layer, hidden layer, and output layer to output vector Y; in the backward feedback process, the output vector Y of the forward feedback is compared with the target vector Y. target The error content between them is used to reversely solve the weight adjustment gradient of the output layer and the hidden layer to form a new numerical model. The specific algorithm flow is as follows:

[0042] Process 1: Build the network structure: Assume that there are n nodes in the input layer (the number of nodes in the input layer is equal to the length of the input vector), p nodes in the hidden layer, and q nodes in the output layer (the number of nodes in the output layer is equal to the length of the output vector);

[0043] Process 2: Let the input vector be X = [x1, x2, ..., x n ], hidden layer vector H = [h1, h2, ..., h p ], the output layer vector Y = [y1, y2, ..., y q ], the expected output vector T = [t1, t2, ..., t q ], the weight between the input layer and the hidden layer is W 1 , B 1 is the bias vector between the two layers: the weight between the hidden layer and the output layer is W 2 , B 2is the bias vector between the two layers;

[0044] Process 3, after initializing the weight W and the bias vector B, the expression for calculating the hidden layer vector from the input layer vector is:

[0045] H=f(W 1 X+B 1 ) Formula 7

[0046] The expression for calculating the output layer vector from the hidden layer vector is:

[0047] Y=f(w 2 H+B 2 ) Formula 8

[0048] Thus, the calculation of the output vector Y of the forward feedback is completed;

[0049] Process 4: Calculate the error function between the expected output vector T and the vector Y, and reversely derive the weight adjustment gradient of the output layer and the hidden layer based on the relationship between the input layer, hidden layer, and output layer described in process 3. The error function is:

[0050]

[0051] The error is back-propagated from the output layer to the hidden layer, and the weight adjustment of each layer is updated using the gradient descent method, that is, W = W-ηΔE p (W, B), where η is the set learning rate, is the step size of each update, and ΔE p (W, B) is the partial derivative of the output deviation under the p-th sample input with respect to the weight of a certain layer, indicating that the parameter is updated once each sample is input;

[0052] Process 5: Based on the above weight adjustment method, calculate the new weight value, implement the backward feedback process of the neural network, and update the weight value;

[0053] Process 6: Calculate the global error and determine whether the network error meets the requirements. When the error reaches the preset accuracy or the number of learning times is greater than the set maximum number, the algorithm ends. Otherwise, select the next learning sample and the corresponding expected output, return to process 3, and enter the next round of learning.

[0054] When constructing a flood prediction model, this application adds factors such as evapotranspiration, air humidity, temperature, and air pressure as inputs based on the BP neural network model to improve the accuracy of water level prediction;

[0055] The flood disaster prediction model is a service module for re-judging flood events based on the flood disaster calculation model. This part provides prediction services during flood disasters based on the flood event preparation and response stage status detected by the flood disaster calculation model, calculates the water level data for the next two days, and then judges the subsequent flood trend based on the stage water level critical value in the flood disaster calculation model. On this basis, the BP neural network water level prediction service is realized to improve the accuracy of stage detection of flood events.

[0056] Preferably, the real-time service framework of the large-scale flood event model is a highly time-sensitive flood service framework that includes real-time acquisition and release of flood data, real-time exploration and calculation of flood status, and detection of flood status changes and active service response.

[0057] The framework is based on sensor monitoring services and event services, constructs sensor access methods and flood active service models, and completes real-time flood disaster information service methods. The framework consists of four parts: flood data basic services, flood subscription service registration center, flood disaster detection, and flood active services.

[0058] The flood data basic service is the driving service of the framework. The sensor streaming data access unit accesses the sensor monitoring data in real time, converts the sensor byte stream into the data stream under the SWE framework, and inserts it into the sensor monitoring service in real time. The sensor monitoring service exposes the accessed flood data records in real time through the interface and publishes the data content in real time.

[0059] The flood subscription service registration center is the flood event subscription interface provided by the framework. Users can define the flood disaster AI model through the subscription service center. The flood disaster calculation model defined by the user in the flood subscription service registration center is encoded in EML and acts on the sensor event service to provide a detection model for flood message production. The defined flood disaster prediction service acts on the flood active service, clarifies the water level prediction model structure, and assists in the re-judgment of future flood disaster information.

[0060] The sensor event service is the core of flood disaster detection. The event server filters the specific monitoring data pulled from the sensor monitoring service by the link layer according to the flood mode subscribed by the registration center, detects the flood disaster status information, publishes the message object, and forms a message queue. The message object Message includes the flood event identifier EventID, the flood event detection time DetectedTime, and the flood event status type EventPhaseType;

[0061] Message=[EventID, DetectedTime, EventPhaseType] Equation 10

[0062] The flood active service automatically converts the flood stage status into flood information service content based on the flood message content generated by flood stage exploration, and is responsible for consuming flood messages. In the active service module, the message queue is detected through changes, and the flood active service type and content are triggered and changed in real time. When the diagnosis is transformed into the preparation stage or the preparation is transformed into the response stage, the scheduled flood forecast and warning service is triggered; when the response is transformed into the recovery stage, the flood status degradation information service is released; when the recovery is transformed into the diagnosis stage, that is, at the end of the recovery stage, flood disaster information calculation services are provided.

[0063] Preferably, the sensor access method with storage and computing separation: data reception, data filtering and data storage are separated from each other, and each component unit is independent of each other and has high reusability. Based on the Storm streaming processing framework, the sensor data access is divided into a data receiving component, a monitoring filtering component and a monitoring storage component. The access framework takes the Storm topology monitoring access unit in the topology layer as the core, converts the byte stream data of the sensor network layer into a data stream that meets the OGCSWE monitoring service standard, and completes the rapid acquisition, parsing, filtering and storage of sensor monitoring data with the support of the SOS service;

[0064] 1) Data receiving component: Under the OGC SWE standard, the data receiving component obtains the sensor eight-tuple metamodel from the SOS service, builds a unified access model in a sensor site-centric manner, and organizes information such as the sensor's geographic spatial location, monitoring time, and monitoring attributes under the sensor site. The unified access model is an information storage unit for sensor attribute monitoring values, and describes sensor monitoring in the form of "sensor site-sensor-monitoring attribute". Sites and sensors, sensors and monitoring attributes are organized in a one-to-many manner. In addition to building a unified sensor access model, the data acquisition component obtains and parses the byte stream transmitted by the sensor from the network layer, and obtains the corresponding sensor data transmission protocol and data expression-related data access metadata from the sensor metamodeling. The structure of the data expression is in the form of an array of [SensorID, PropertyID, DataPosition, DataLength, DataRatio], where SensorID refers to the current sensor ID, PropertyID is the sensor monitoring attribute ID, DataPosition is the data position, DataLength is the data length, and DataRatio is the ratio of the monitoring data to the target unit;

[0065] 2) Monitoring and filtering components: Sensor monitoring is an expression of geographical phenomena related to space and time. The monitoring and filtering components solve the monitoring scene limitations encountered during the sensor monitoring access process, and filter the time, space and monitoring values. The filtering module supports multi-condition filtering based on the OGC Filter standard. It connects each filtering unit with AND logic operations, and supports monitoring value filtering in spatiotemporal scenarios. Spatial filtering is filtering for sensor spatial scenes, and filtering for spatial points, including methods such as inclusion, non-inclusion and rectangular query. Time filtering is filtering for monitoring time scenes, filtering time point information, including operations within a certain time range, before a certain time and after a certain time. Attribute filtering is filtering for abnormal values, and supports basic comparison filtering operations such as greater than, less than and equal to.

[0066] 3) Monitoring storage component: After the sensor monitoring data is collected, parsed, and filtered, its data format and expression meet the requirements of the corresponding application scenario. At this time, the monitoring storage component converts its data structure into a standard monitoring data model structure encoded in O&M, and constructs a monitoring data service insertion request to realize the rapid insertion process of sensor monitoring data;

[0067] The sensor data receiving component, monitoring filtering component and monitoring storage component are the basic units for sensor monitoring data access. The Storm streaming data processing framework integrates the three components on this basis, constructing the data receiving component with the Spout unit, and the monitoring filtering component and monitoring storage component with the Blot unit, forming a sensor monitoring access unit that meets the Storm platform standard. The topology standardizes the sensor access process and ensures that the sensor access has the possibility of real-time change. The access process uses the topology template to build the topology in the XMLSchema standard. The configuration mode includes four parts: topology identifier TopologyID, sensor access component Receive, sensor filtering component Filter and sensor monitoring storage component Store.

[0068] After the topology node is uploaded and started, the sensor access unit enters the running state. The state management is controlled by the RestflAPI service provided by StormUI to turn on and off the sensor access topology unit.

[0069] Preferably, the flood active service: provides three active service contents: early warning, water level prediction and flood calculation. The flood early warning, alarm and downgrade release service is a continuous phase service acting on the active service module, corresponding to the water level prediction service in the flood active service; the flood water level prediction service starts from the flood preparation stage and ends automatically after the flood recovery stage. In the preparation stage, the system provides flood early warning and prediction services; in the response stage, it also provides flood early warning and flood prediction services; in the recovery stage, it provides flood downgrade notification, flood prediction and flood calculation services;

[0070] 1) Flood warning service: Publish flood status information in real time in the form of emails. The system automatically edits flood event warning information content based on the current flood status, calls the email service through the email server, notifies users of flood event status information, and generates flood warning records;

[0071] 2) Flood prediction service: Based on the current water level of the upstream water level station and the meteorological data information in the basin, the water level data of the target monitoring station in the next two days is predicted in real time. The steps are as follows: Step 1: Enter the current event ID and time, obtain the parameters of the flood disaster AI large model, and obtain the prediction model parameter configuration; Step 2: According to the prediction model parameter configuration, obtain flood monitoring data, and align the data content to form the prediction model training data and prediction input data; Step 3: Normalize the training data, initialize the BP neural network model parameters, and train the neural network model; Step 4: Normalize the prediction data, and use the trained neural network model to predict the water level to obtain the normalized prediction result; Step 5: Based on the normalized prediction result, perform denormalization, obtain the water level prediction result, and store the prediction result.

[0072] Preferably, flood calculation service: flood disaster object refers to the complete flood event object detected in the flood disaster information service, including the start and end time of the flood event, the start time of the flood diagnosis phase, the start time of the preparation phase, the start time of the response phase, and the start and end time of the recovery phase, describing the information of each stage of the entire flood event process. The flood calculation service calculates flood disaster information, automatically generates, stores, and manages flood event process objects, and calculates the maximum water level information during the flood occurrence stage, so as to prepare data for subsequent flood event analysis and research;

[0073] Flood disaster calculation service process steps: Step a: According to the flood event ID and the current time point, obtain the historical message time series set from the end of the previous event to the current event, and initialize the flood disaster object; Step b: According to the message type classification, the current message is divided into four stages: diagnosis, preparation, response and recovery; Step c: If the current message type is the diagnosis stage, determine whether the message type has changed. If it has changed, update the flood event status, otherwise update the preparation stage and event start time of the flood disaster object; and determine whether the previous message type is the recovery stage. If so, set the flood disaster object recovery stage and event end time to store the flood disaster object and calculate the maximum water level. Otherwise, update the preparation stage and event start time of the flood disaster object; If the current message type is the preparation, response or recovery stage, update the current event status and set the start time of the corresponding stage, check whether the message set is traversed, if so, throw an exception, otherwise, return to step b.

[0074] Preferably, the business process of the real-time service system for the large model of flood events includes four business parts: user subscription to flood events, large model exploration and service of flood events, sensor access management, and flood event management and viewing;

[0075] In the user subscription service, users configure the flood event subscription model, select the monitoring data set related to the flood event from the sensor monitoring service, and then configure the flood disaster calculation model parameters, flood disaster prediction model and other parameters. After that, the flood event subscription model is registered in the sensor event service and managed uniformly in the flood database server;

[0076] In the business of large-scale model exploration and service of flood events, hydrological and meteorological sensors, under the premise of modeling registration, access sensor monitoring data in real time through sensor access units, and publish monitoring data in real time through sensor monitoring services. After the monitoring data is uniformly released by SOS, it flows into the sensor event exploration server, explores each flood event stage in real time, triggers the service modules of each stage of flood events, and actively sends service results to users in the form of emails, completing the active exploration and service of flood events. The exploration results and stage service results will also be recorded in the flood database to serve the subsequent flood management and information display business;

[0077] In the sensor access management service, users can view, delete, and add sensor access units to achieve visual management of access sensors. In the flood management and viewing service, users can obtain flood history content through sensor monitoring services, sensor event services, and flood databases, and actively view flood sensor information, flood sensor monitoring data information, flood event subscription information, flood event stage status information, and flood stage service information.

[0078] Based on the above business processes, the system carries out overall framework design, functional design and database design, and clarifies the main functions, task processes and data structures of the real-time service of the large model of flood events.

[0079] Preferably, the system backend service: adopts a four-layer hierarchical system, dividing the backend into a user interaction layer, a business logic layer, an intermediate service layer, and a database layer;

[0080] User interaction layer: provides corresponding interface support for the operation units in the system interface, and directly relies on the business logic layer of the lower layer, that is, the business logic layer provides specific algorithm implementation for the user interaction layer, and the front-end interface calls the user interaction layer interface to complete the business process; the operation interface of this layer includes login interface, map operation interface, flood management interface, flood sensor access unit management interface and flood subscription model parameter configuration interface, and the active service interface directly sends emails to users through email services to notify the stage status of flood events and water level prediction results; with the support of the interface, users can log in to this system and upload sensor access units by themselves to obtain data in specific scenarios; at the same time, users subscribe to flood events in a certain geographical scenario and view the corresponding flood disaster information through the flood event management unit;

[0081] Business logic layer: responsible for calling the service content of the middle service layer, and after parsing or encoding the data, integrating the business process according to business needs, and providing the required data objects and algorithm results to the user interaction layer and the middle service layer. The business logic includes flood event subscription coding, OGCSWE standard parsing, login logic judgment, sensor monitoring access and flood stage services;

[0082] Middle service layer: abstracts the basic service units in the system into SOA services, bringing business reuse. At the same time, each service unit in the middle service layer expands its own performance. Sensor monitoring uses stand-alone services based on PostgreSQL storage or distributed services based on MongoDB storage. Based on the principle of active service module, it can actively detect flood events and trigger flood service content in the corresponding stage. The sensor access unit runs in real time on the Storm platform, registers the flood data content perceived by the flood sensor in real time to the sensor monitoring service, and the Feeder link service regularly pulls the flood data content required by the sensor event detection service and injects it into the sensor event server in real time. The sensor event server can actively detect flood events subscribed by users, and finally actively publish flood event status information through the network notification service to complete automated flood disaster detection.

[0083] Database layer: stores all data content in the real-time service system for flood disaster information, among which the SOS database stores the flood sensor model and its monitoring data content; the Feeder database stores the linked SOS data source and the sensor information that needs to be connected to SES; the SES subscription management library manages event subscription information in the form of files; the flood database stores all object information involved in the business logic service layer of this system.

[0084] Compared with the prior art, the innovations and advantages of this application are:

[0085] (1) After analyzing the current flood detection system, this application proposes a real-time service framework for flood disaster information based on the OGCSWE sensor network integration framework. Based on the flood disaster AI big model, it detects the diagnosis, preparation, response and recovery stages of flood events, and automatically triggers flood prediction services, warning services and computing services in real time, thereby providing active flood disaster information services with high granularity: 1) A flood disaster AI big model is designed, combining flood disaster detection with BP neural network water level prediction content supported by multiple meteorological data to predict water levels during floods; 2) A real-time service framework for flood event big model based on the SWE framework is designed, which clarifies the access method of sensor streaming data, flood Flood event detection mechanism and active service method; 3) A real-time service system for a large model of flood events based on the SWE framework was realized; 4) In flood event detection, this system can detect flood events in advance and automatically form flood disaster objects, and the detection results are more detailed than the current system; in water level prediction, after the introduction of meteorological data such as evapotranspiration and relative humidity, the water level prediction results are significantly improved; in terms of active service real-time, the system can complete the flood event stage detection, service triggering and service release process within 2.5 minutes, and it has good real-time performance; in terms of system scalability, based on the SWE framework standard and the system's internal service interface standard, it provides external scalability support, thus ensuring the system's good scalability.

[0086] (2) In order to solve the problem of information islands and service barriers caused by system-specific standards in flood event detection, improve the granularity of flood event detection, and enrich flood system services, this application builds a real-time service system for a large model of flood events based on the SWE framework, from data access to flood sensors to detection and processing of flood events based on OGC Based on the SWE sensor network integration framework, a multi-source heterogeneous flood sensor registration and real-time access and storage of monitoring data are established, and the diagnosis, preparation, response and recovery stages of flood events are explored under unified standards, and flood warnings, hydrological forecasts and flood event calculation services are issued in real time; specifically, a flood disaster AI model, a flood event model real-time service framework based on the SWE framework, and a flood event model real-time service system are established. They have a real-time system foundation, which can solve the problems of timely warning, release and prediction of flood events, and provide favorable conditions for relevant departments to make quick decisions and respond effectively; the BP neural network is used to predict water levels. In the prediction method based on precipitation and upstream water levels, new meteorological factors such as temperature, air pressure, and atmospheric humidity are added, the prediction accuracy is improved, and high-efficiency, high-availability and high-granularity flood information services are provided.

[0087] (3) The real-time service system of the flood event big model based on the SWE framework proposed in this application, under the OGCSWE sensor network integration framework, uses the big AI model of the flood event process to express the past, present and future changing characteristics of floods, explore the development process of flood events, and divide it into diagnosis, preparation, response and recovery stages. Compared with the traditional flood event detection and release system, this system not only has good real-time performance, but also can solve the problems of data sharing, unified processing and detection of multi-source data, and real-time active service. Based on the big model of flood event detection, this system improves the granularity of flood event services, enriches the services at all stages of floods, provides users with flood warning, water level prediction and event disaster calculation services, provides real-time and active flood status analysis and decision support for relevant disaster prevention and control departments, prepares for reducing casualties and property losses caused by flood disasters, and provides decision-making departments with more detailed flood disaster impact assessment and command decision support. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is the overall framework diagram of the AI ​​model for flood disasters.

[0089] Figure 2 This is a real-time service framework diagram for a large model of flood events.

[0090] Figure 3 It is a schematic diagram of the sensor data access framework.

[0091] Figure 4 It is a flow chart of flood disaster calculation service.

[0092] Figure 5 It is the business process diagram of the real-time service system for flood disaster information.

[0093] Figure 6 This is a conceptual model diagram of the flood event process information real-time service system database.

[0094] Figure 7 It is a service delay diagram of the flood event process information real-time service system based on the SWE framework. DETAILED DESCRIPTION

[0095] Below, in conjunction with the accompanying drawings, the technical solution of the real-time management system of the AI ​​large model of flood disasters in the water conservancy sensor network provided by this application is further described so that technical personnel in this field can better understand this application and implement it.

[0096] With the continuous development of sensor networks and the Internet of Things, reliability and real-time features have become more prominent. More and more hydrological monitoring departments have begun to introduce sensor equipment to automatically obtain hydrological monitoring data in geographic space in real time. Based on this, the real-time information system of flood events has gradually become an important tool for flood monitoring and intelligent decision-making. However, the current flood information system is mostly based on special standards, which makes it difficult to share flood information and actively provide flood services. In addition, the granularity is poor in the process of flood event detection, and the flood content is only clarified by whether it occurs or not, which has an important impact on the management of flood information and further flood information services.

[0097] After analyzing the current flood detection system, this application proposes a real-time service framework for flood disaster information based on the OGCSWE sensor network integration framework. Based on the AI ​​big model of flood disasters, it detects the diagnosis, preparation, response and recovery stages of flood events, and automatically triggers flood prediction services, warning services and computing services in real time, thereby providing active flood disaster information services with high granularity. (1) Design a big AI model of flood disasters, combine flood disaster detection with BP neural network water level prediction content supported by multiple meteorological data, divide flood events into four process stages: diagnosis, preparation, response and recovery, and perform water level prediction during the flood process; (2) Design a real-time service framework for flood event big model based on the SWE framework, clarify the access method of sensor streaming data, flood event detection mechanism and active service method; (3) According to the designed content, realize the real-time service system of flood event big model based on the SWE framework; (4) Select the Hanjiang sub-basin as the experimental area, simulate and detect flood events, and publish flood services in real time. The results show that in terms of flood event detection, the system can detect flood events in advance and automatically form flood disaster objects, and the detection results are more precise than the current system; in terms of water level prediction, the water level prediction results are significantly improved after the introduction of meteorological data such as evapotranspiration and relative humidity; in terms of active service real-time, the system can complete the flood event stage detection, service triggering and service release process within 2.5 minutes, and it has good real-time performance; in terms of system scalability, based on the SWE framework standard and the system's internal service interface standard, it provides external scalability support, thus ensuring the system's good scalability.

[0098] 1. Flood Disaster AI Model

[0099] (I) Large Model Architecture

[0100] Based on the water level monitoring method, flood event information is detected in real time. The water level critical value method is used as the basis for dividing flood disasters, and flood disasters are divided into four stages: diagnosis, preparation, response and recovery. In the use value division process, based on the single frequency value method and the multi-frequency critical value method, in the flood detection process, it is ensured that users can judge the flood status based on a single water level data, and eliminate the erroneous flood warning caused by water level fluctuations based on the water level data frequency calculation within a certain time range; in the flood AI large model of this application, on the basis of the single value method, the flood event detection model under the multi-frequency calculation of water level data is adopted as the flood disaster calculation model. Based on the spatial distribution of water level sensors and the different monitoring frequencies, for different water level sensors, the parameters and critical value information of the calculation model are established according to the calculation analysis of the historical flood occurrence process or the reference to the warning water level and guaranteed water level information content in the river water level.

[0101] The model computationally represents historical and current flood status information, describes flood disasters in a four-stage manner, provides early warning in the preparation stage, and predicts the future trend of water level data at the target site based on historical monitoring data from hydrological, meteorological and other sensors during the flood preparation and response stages.

[0102] The AI ​​big model for flood disasters includes two parts: the flood disaster calculation model and the flood disaster prediction model. The flood disaster calculation model divides flood stages and clarifies the historical and current flood status, while the prediction model represents future flood trends. The AI ​​big model for flood disasters refines flood stages and represents the past, present and future flood status.

[0103] (II) Model composition

[0104] The model is a rule for flood event detection. It defines the characteristics of factors at each stage of flood events based on the direct characterization factors of the current flood situation, including water level, flow and other information, and defines the flood event information process model based on the characteristic pattern of water level changes. Figure 1 The AI ​​model of flood disasters consists of two parts. The first part is a characterization model that calculates whether the monitoring frequency of water level monitoring values ​​in a given interval is greater than or equal to the critical value within the past to the current time range. As a flood disaster calculation model, based on the water level as a direct characterization factor of floods, flood events are divided into four stages: diagnosis, preparation, response and recovery, expressing the flood status within the time range from now to the past TW; the second part is a flood disaster prediction model that predicts water levels. Based on historical water levels and meteorological data, under the premise of factor normalization, the prediction model is trained based on the BP neural network to realize water level prediction. Based on the flood disaster calculation model in the first part, it provides future water level prediction function when the flood disaster is in the preparation response and recovery stages. Based on the water level prediction, the future flood trend is analyzed.

[0105] Define W to represent water level data, U(N) to represent the critical water level interval of stage N, where N∈[A, B, C, D], A represents the diagnosis stage, B represents the preparation stage, C represents the response stage, and D represents the recovery stage. TW(N) represents the time window of stage N, and Fre(TW(N)) refers to the event frequency of the water level WEU(N) of stage N within the TW(N) time window. F(N) represents the critical value of the calculated frequency of stage N. The algorithm steps are as follows:

[0106] Step 1: Input water level monitoring data; Step 2: When judging the water level W∈U(N), the value space P of N; Step 3: Within the value range of P, calculate the frequency Fre(N) of the event type in the TW(N) time window, and compare whether it is greater than F(N), where N∈P. If the calculated frequency is greater than the given value, the object is recorded, and the value space E of N is further obtained; Step 4: For each value in the value space E of N, loop; if the value is A, output directly; if the value is B, determine whether the state of the previous flood event is A, if so, output stage B, otherwise discard; if the value is C, Then determine whether the status of the previous flood event is B. If so, output stage C, otherwise discard it: If the value is D, then determine whether the status of the previous flood event is C. If so, output D, otherwise discard it; Step 5: According to the flood stage output in the fourth step, determine whether the flood event is in the preparation, response and recovery stages; If so, enter the sixth step to provide flood disaster prediction services; Step 6: Select historical water levels and meteorological data within the TW1 time range, normalize them to form training data, and train the BP neural network model; According to the current monitoring data within the TW2 time range, construct prediction data, predict water levels, and obtain prediction results R.

[0107] 1. Flood disaster calculation model

[0108] It is divided into diagnosis, preparation, response and recovery phase filtering conditions. The exploration of each phase (except the diagnosis phase) must meet two basic filtering conditions: hydrological monitoring data filtering and event phase dependency filtering;

[0109] Among them, each flood event stage needs to meet the basic hydrological monitoring data filtering conditions:

[0110] Condi=water level∈U&Fre(TW)≥F Formula 1

[0111] The frequency Fre of the water level monitoring data shown in Formula 1 appearing in a given water level interval U within a certain time TW is greater than or equal to the given frequency F;

[0112] Formula 2 is the dependent filtering condition of the flood event stage, which is the key to ensure the mutual dependence of each stage of the flood event. The definition of the flood event stage must strictly meet the sequential dependence relationship of A->B->C->D. The definition of the current stage CurrentPhase must be based on meeting the filtering conditions of the hydrological monitoring data and strictly depend on the previous exploration results;

[0113] CurrentPhase=Condi after astPhase, astPhase∈[A, B, C] Equation 2

[0114] The filter condition A that the monitoring data in the diagnosis phase should meet is:

[0115] A=water level<W1&Fre(TW=T1)≥F1 Formula 3

[0116] The filter condition B that the monitoring data in the preparation phase should meet is:

[0117] BCondi=W1≤water level <W2&Fre(TW=T2)≥F2

[0118] B=Bcondi after A Formula 4

[0119] The filtering condition C that the monitoring data in the response phase should meet is:

[0120] Ccondi=water level≥W2&Fre(TW=T3)≥F3

[0121] C=Ccondi after B Formula 5

[0122] The filtering condition D that the monitoring data in the recovery phase should meet is:

[0123] Dcondi=water level<W3&Fre(TW=T4)≥F4

[0124] D=Ccondi after C Formula 6

[0125] W1, W2, and W3 are the water level values at different stages, and satisfy W1 < W2 and W3 < W2; TW, short for TimeWindow, refers to the time window, and T1, T2, T3, and T4 are the critical values of different time windows respectively, and satisfy T1 ≥ T2 ≥ T3 ≥ T4; Fre(TW) refers to the number of times the water level values that meet the conditions appear within the time window TW, and F1, F2, F3, and F4 are different numerical values. By restricting the relative sizes of the water level values and the length of the time window, the confusion between different stages of flood events is avoided, ensuring the uniqueness of the stage where the flood event is located. In addition, the diagnosis, preparation, response, and recovery stages are in strict sequence. For the three conditions B, C, and D, in addition to judging that they meet the current basic conditions Bcondi, Ccondi, and DCondi, they also need to meet the change trend of the stage evolution of A->B->C->D, that is, the preparation stage must be after the diagnosis stage, the response stage must be after the preparation stage, and the recovery stage must be after the response stage in turn.

[0126] 2. Flood Disaster Prediction Model

[0127] The calculation model characterizes the change characteristics of hydrological monitoring data from the past to the present, while the prediction model conducts hydrological prediction of flood events based on the flood event status information extracted after preliminary filtering by the calculation model, improving the accuracy and timeliness of re-judging the occurrence time and status of flood events.

[0128] A short-term water level prediction model is constructed based on the BP neural network, which consists of two major parts: forward feedback and backward feedback. During the forward feedback process, the input vector X passes through the input layer, hidden layer, and outputs the vector Y at the output layer; during the backward feedback process, according to the error content between the forward feedback output vector Y and the target vector Y target The weight adjustment gradients of the output layer and the hidden layer are calculated in reverse to form a new numerical model. The specific algorithm flow is as follows:

[0129] Process 1: Construct the network structure: Assume that the input layer has n nodes (the number of input layer nodes is equal to the length of the input vector), the hidden layer has p nodes, and the output layer has q nodes (the number of output layer nodes is equal to the length of the output vector);

[0130] Process 2: Assume the input vector is X = [x1, x2,..., x n , the hidden layer vector H = [h1, h2,..., h p , the output layer vector Y = [y1, y2,..., y q , the expected output vector T = [t1, t2,..., t q , the weight between the input layer and the hidden layer is W 1 , B 1 is the bias vector between the two layers: the weight between the hidden layer and the output layer is W2 , B 2 is the bias vector between the two layers;

[0131] Process 3, after initializing the weight W and the bias vector B, the expression for calculating the hidden layer vector from the input layer vector is:

[0132] H=f(W 1 X+B 1 ) Formula 7

[0133] The expression for calculating the output layer vector from the hidden layer vector is:

[0134] Y=f(W 2 H+B 2 ) Formula 8

[0135] Thus, the calculation of the output vector Y of the forward feedback is completed;

[0136] Process 4: Calculate the error function between the expected output vector T and the vector Y, and reversely derive the weight adjustment gradient of the output layer and the hidden layer based on the relationship between the input layer, hidden layer, and output layer described in process 3. The error function is:

[0137]

[0138] The error is back-propagated from the output layer to the hidden layer, and the weight adjustment of each layer is updated using the gradient descent method, that is, W = W-ηΔE p (W, B), where η is the set learning rate, is the step size of each update, and ΔE p (W, B) is the partial derivative of the output deviation under the p-th sample input with respect to the weight of a certain layer, indicating that the parameter is updated once each sample is input;

[0139] Process 5: Based on the above weight adjustment method, calculate the new weight value, implement the backward feedback process of the neural network, and update the weight value;

[0140] Process 6: Calculate the global error and determine whether the network error meets the requirements. When the error reaches the preset accuracy or the number of learning times is greater than the set maximum number, the algorithm ends. Otherwise, select the next learning sample and the corresponding expected output, return to process 3, and enter the next round of learning.

[0141] When constructing a flood prediction model, this application adds factors such as evapotranspiration, air humidity, temperature, and air pressure as inputs based on the BP neural network model to improve the accuracy of water level prediction;

[0142] The flood disaster prediction model is a service module for re-judging flood events based on the flood disaster calculation model. This part provides prediction services during flood disasters based on the flood event preparation and response stage status detected by the flood disaster calculation model, calculates the water level data for the next two days, and then judges the subsequent flood trend based on the stage water level critical value in the flood disaster calculation model. On this basis, the BP neural network water level prediction service is realized to improve the accuracy of stage detection of flood events.

[0143] 2. Real-time service framework for large-scale flood event models

[0144] (I) Service framework structure

[0145] The real-time status change of flood immediately triggers the corresponding service. The framework focuses more on the change detection of flood status and the active response of services. The change detection of flood event status relies on the real-time exploration and acquisition of flood event status data stream, while the exploration of flood status depends on the real-time acquisition of hydrological, meteorological and other sensor monitoring data. The real-time service framework of the large model of flood events is a highly time-sensitive flood service framework, from the real-time acquisition and release of flood data to the real-time exploration and calculation of flood status, to the detection of flood status changes and the active response of services.

[0146] like Figure 2 This is a diagram of the real-time service framework for the large model of flood events. The framework is based on sensor monitoring services and event services, builds sensor access methods and flood active service models, and completes the real-time service method for flood disaster information. The framework consists of four parts: flood data basic services, flood subscription service registration center, flood disaster detection, and flood active services.

[0147] The flood data basic service is the driving service of the framework. The sensor streaming data access unit accesses the sensor monitoring data in real time, converts the sensor byte stream into the data stream under the SWE framework, and inserts it into the sensor monitoring service in real time. The sensor monitoring service exposes the accessed flood data records in real time through the interface and publishes the data content in real time.

[0148] The flood subscription service registration center is the flood event subscription interface provided by the framework. Users can define the flood disaster AI model through the subscription service center. The flood disaster calculation model defined by the user in the flood subscription service registration center is encoded in EML and acts on the sensor event service to provide a detection model for flood message production. The defined flood disaster prediction service acts on the flood active service, clarifies the water level prediction model structure, and assists in the re-judgment of future flood disaster information.

[0149] The sensor event service is the core of flood disaster detection. The event server filters the specific monitoring data pulled from the sensor monitoring service by the link layer according to the flood mode subscribed by the registration center, detects the flood disaster status information, publishes the message object, and forms a message queue. The message object Message includes the flood event identifier EventID, the flood event detection time DetectedTime, and the flood event status type EventPhaseType;

[0150] Message=[EventID, DetectedTime, EventPhaseType] Equation 10

[0151] The flood active service automatically converts the flood stage status into flood information service content based on the flood message content generated by flood stage exploration, and is responsible for consuming flood messages. In the active service module, the message queue is detected through changes, and the flood active service type and content are triggered and changed in real time. When the diagnosis is transformed into the preparation stage or the preparation is transformed into the response stage, the scheduled flood forecast and warning service is triggered; when the response is transformed into the recovery stage, the flood status degradation information service is released; when the recovery is transformed into the diagnosis stage, that is, at the end of the recovery stage, flood disaster information calculation services are provided.

[0152] 2. Sensor access method with storage and computing separation

[0153] Data reception, data filtering, and data storage are separated from each other, and each component unit is independent of each other, with high reusability. Based on the Storm streaming processing framework, sensor data access is divided into data reception components, monitoring filtering components, and monitoring storage components. Figure 3 ,The access framework takes the Storm topology monitoring access unit in the topology layer as the core, ,which converts the byte stream data of the sensor network layer into a data stream ,that meets the OGCSWE monitoring service standard, and ,completes the rapid acquisition, parsing, filtering and storage of sensor monitoring ,data with the support of the SOS service;

[0154] 1) Data receiving component: Under the OGC SWE standard, the data receiving component obtains the sensor eight-tuple metamodel from the SOS service, builds a unified access model in a sensor site-centric manner, and organizes information such as the sensor's geographic spatial location, monitoring time, and monitoring attributes under the sensor site. The unified access model is an information storage unit for sensor attribute monitoring values, and describes sensor monitoring in the form of "sensor site-sensor-monitoring attribute". Sites and sensors, sensors and monitoring attributes are organized in a one-to-many manner. In addition to building a unified sensor access model, the data acquisition component obtains and parses the byte stream transmitted by the sensor from the network layer, and obtains the corresponding sensor data transmission protocol and data expression-related data access metadata from the sensor metamodeling. The structure of the data expression is in the form of an array of [SensorID, PropertyID, DataPosition, DataLength, DataRatio], where SensorID refers to the current sensor ID, PropertyID is the sensor monitoring attribute ID, DataPosition is the data position, DataLength is the data length, and DataRatio is the ratio of the monitoring data to the target unit;

[0155] 2) Monitoring and filtering components: Sensor monitoring is an expression of geographical phenomena related to space and time. The monitoring and filtering components solve the monitoring scene limitations encountered during the sensor monitoring access process, and filter the time, space and monitoring values. The filtering module supports multi-condition filtering based on the OGC Filter standard. It connects each filtering unit with AND logic operations, and supports monitoring value filtering in spatiotemporal scenarios. Spatial filtering is filtering for sensor spatial scenes, and filtering for spatial points, including methods such as inclusion, non-inclusion and rectangular query. Time filtering is filtering for monitoring time scenes, filtering time point information, including operations within a certain time range, before a certain time and after a certain time. Attribute filtering is filtering for abnormal values, and supports basic comparison filtering operations such as greater than, less than and equal to.

[0156] 3) Monitoring storage component: After the sensor monitoring data is collected, parsed, and filtered, its data format and expression meet the requirements of the corresponding application scenario. At this time, the monitoring storage component converts its data structure into a standard monitoring data model structure encoded in O&M, and constructs a monitoring data service insertion request to realize the rapid insertion process of sensor monitoring data;

[0157] The sensor data receiving component, monitoring filtering component and monitoring storage component are the basic units for sensor monitoring data access. The Storm streaming data processing framework integrates the three components on this basis, constructing the data receiving component with the Spout unit, and the monitoring filtering component and monitoring storage component with the Blot unit, forming a sensor monitoring access unit that meets the Storm platform standard. The topology standardizes the sensor access process and ensures that the sensor access has the possibility of real-time change. The access process uses the topology template to build the topology in the XMLSchema standard. The configuration mode includes four parts: topology identifier TopologyID, sensor access component Receive, sensor filtering component Filter and sensor monitoring storage component Store.

[0158] After the topology node is uploaded and started, the sensor access unit enters the running state. The state management is controlled by the RestflAPI service provided by StormUI to turn on and off the sensor access topology unit.

[0159] (III) Flood disaster detection services

[0160] It is a derivative service of the sensor event service. It uses the flood disaster calculation model to calibrate the sensor event service so that it has the characteristics of flood disaster detection. It builds a flood subscription model and registers it with the sensor event service to filter flood basic data, detect flood disasters, and publish flood disaster status through the WNS service.

[0161] The flood disaster detection service relies on the sensor monitoring service, which is a storage and publishing unit for heterogeneous flood data. Heterogeneous hydrological and meteorological data are expressed in the same O&M code and published through SOS standard packaging, so that flood sensor monitoring data can be exposed in real time.

[0162] The sensor monitoring service and the sensor event service are both independent services. They do not have time dynamics and therefore need to be linked to each other by the link layer and driven in real time to form an active service process. The SOSSESFeeder service pulls flood monitoring data in real time through the GetObservation interface of SOS and registers it in SES in real time through the Notify interface of SES to ensure the initiative of flood disaster detection services. In addition, SOSSESFeeder filters out unnecessary sensor data streams in SES to ensure the validity of access data.

[0163] The flood disaster detection service is composed of four modules: sensor monitoring service, SOSSESFeeder link service, sensor event service and network notification service. The service completes the real-time detection of flood event status under the SWE framework. After the hydrological monitoring filtering and stage dependency filtering in the flood calculation model, the solution space of the flood event status is screened out and published in real time to ensure the real-time and proactive nature of the real-time flood disaster information service.

[0164] (IV) Active flood services

[0165] It provides three kinds of active services: early warning, water level prediction and flood calculation. The flood early warning, alarm and downgrade release service is a continuous phase service acting on the active service module, corresponding to the water level prediction service in the flood active service. The flood water level prediction service starts from the flood preparation stage and ends automatically after the flood recovery stage. In the preparation stage, the system provides flood early warning and prediction services; in the response stage, it also provides flood early warning and flood prediction services; in the recovery stage, it provides flood downgrade notification, flood prediction and flood calculation services;

[0166] 1) Flood warning service: Publish flood status information in real time in the form of emails. The system automatically edits flood event warning information content based on the current flood status, calls the email service through the email server, notifies users of flood event status information, and generates flood warning records;

[0167] 2) Flood prediction service: Based on the current water level of the upstream water level station and the meteorological data information in the basin, the water level data of the target monitoring station in the next two days is predicted in real time. The steps are as follows: Step 1: Enter the current event ID and time, obtain the parameters of the flood disaster AI large model, and obtain the prediction model parameter configuration; Step 2: According to the prediction model parameter configuration, obtain flood monitoring data, and align the data content to form the prediction model training data and prediction input data; Step 3: Normalize the training data, initialize the BP neural network model parameters, and train the neural network model; Step 4: Normalize the prediction data, and use the trained neural network model to predict the water level to obtain the normalized prediction result; Step 5: Based on the normalized prediction result, perform denormalization, obtain the water level prediction result, and store the prediction result.

[0168] 3) Flood calculation service: Flood disaster objects refer to the complete flood event objects detected in the flood disaster information service, including the start and end time of the flood event, the start time of the flood diagnosis phase, the start time of the preparation phase, the start time of the response phase, and the start and end time of the recovery phase, describing the information of each stage of the entire flood event process. The flood calculation service calculates flood disaster information, automatically generates, stores, and manages flood event process objects, and calculates the maximum water level information during the flood occurrence stage, so as to prepare data for subsequent flood event analysis and research;

[0169] Figure 4 The flow chart of flood disaster calculation service is as follows:

[0170] Step a: According to the flood event ID and the current time point, obtain the historical message time series set from the end of the previous event to the current event, and initialize the flood disaster object; Step b: According to the message type classification, the current message is divided into four stages: diagnosis, preparation, response and recovery; Step c: If the current message type is the diagnosis stage, determine whether the message type has changed. If it has changed, update the flood event status, otherwise update the preparation stage and event start time of the flood disaster object; and determine whether the previous message type is the recovery stage. If so, set the flood disaster object recovery stage and event end time to store the flood disaster object and calculate the maximum water level. Otherwise, update the preparation stage and event start time of the flood disaster object. If the current message type is the preparation, response or recovery stage, update the current event status and set the start time of the corresponding stage, check whether the message set traversal is completed, if so, throw an exception, otherwise, return to step b.

[0171] 3. Design of real-time service system for large-scale flood event model

[0172] (I) System business process

[0173] like Figure 5 The business process of the real-time service system for flood disaster information includes four business parts: user subscription to flood events, large-scale model exploration and service of flood events, sensor access management, and flood event management;

[0174] In the user subscription service, users configure the flood event subscription model, select the monitoring data set related to the flood event from the sensor monitoring service, and then configure the flood disaster calculation model parameters, flood disaster prediction model and other parameters. After that, the flood event subscription model is registered in the sensor event service and managed uniformly in the flood database server;

[0175] In the business of large-scale model exploration and service of flood events, hydrological and meteorological sensors, under the premise of modeling registration, access sensor monitoring data in real time through sensor access units, and publish monitoring data in real time through sensor monitoring services. After the monitoring data is uniformly released by SOS, it flows into the sensor event exploration server, explores each flood event stage in real time, triggers the service modules of each stage of flood events, and actively sends service results to users in the form of emails, completing the active exploration and service of flood events. The exploration results and stage service results will also be recorded in the flood database to serve the subsequent flood management and information display business;

[0176] In the sensor access management service, users can view, delete, and add sensor access units to achieve visual management of access sensors. In the flood management and viewing service, users can obtain flood history content through sensor monitoring services, sensor event services, and flood databases, and actively view flood sensor information, flood sensor monitoring data information, flood event subscription information, flood event stage status information, and flood stage service information.

[0177] Based on the above business processes, the system carries out overall framework design, functional design and database design, and clarifies the main functions, task processes and data structures of the real-time service of the large model of flood events.

[0178] 2. System overall framework

[0179] 1. System front-end interface

[0180] It is divided into user interaction interface, system information display interface and development team information interface. The user interaction interface includes system operation modules such as login, home page, sensor access management, flood event subscription and flood event management. The system information display interface displays the main information of the current page, including the selection of flood sensor data set displayed in the flood event subscription page, and the configuration of flood disaster calculation and prediction model parameters. The development team information interface indicates the basic information of the development team so that users can contact the developers.

[0181] The front-end page adopts a single-page design concept. Each user interaction unit corresponds to a different system information display interface. Users can achieve seamless page jumps during system use, improving the user's system smoothness experience.

[0182] 2. System background services

[0183] Adopting a four-layer hierarchical system, the backend is divided into the user interaction layer, business logic layer, intermediate service layer and database layer;

[0184] User interaction layer: provides corresponding interface support for the operation units in the system interface, and directly relies on the business logic layer of the lower layer, that is, the business logic layer provides specific algorithm implementation for the user interaction layer, and the front-end interface calls the user interaction layer interface to complete the business process; the operation interface of this layer includes login interface, map operation interface, flood management interface, flood sensor access unit management interface and flood subscription model parameter configuration interface, and the active service interface directly sends emails to users through email services to notify the stage status of flood events and water level prediction results; with the support of the interface, users can log in to this system and upload sensor access units by themselves to obtain data in specific scenarios; at the same time, users subscribe to flood events in a certain geographical scenario and view the corresponding flood disaster information through the flood event management unit;

[0185] Business logic layer: responsible for calling the service content of the middle service layer, and after parsing or encoding the data, integrating the business process according to business needs, and providing the required data objects and algorithm results to the user interaction layer and the middle service layer. The business logic includes flood event subscription coding, OGCSWE standard parsing, login logic judgment, sensor monitoring access and flood stage services;

[0186] Middle service layer: abstracts the basic service units in the system into SOA services, bringing business reuse. At the same time, each service unit in the middle service layer expands its own performance. Sensor monitoring uses stand-alone services based on PostgreSQL storage or distributed services based on MongoDB storage. Based on the principle of active service modules, it can actively detect flood events and trigger flood service content at the corresponding stage. The sensor access unit runs in real time on the Storm platform, registers the flood data content perceived by the flood sensor in real time to the sensor monitoring service, and the Feeder link service regularly pulls the flood data content required by the sensor event detection service and injects it into the sensor event server in real time. The sensor event server can actively detect flood events subscribed by users, and finally actively publish flood event status information through the network notification service to complete automated flood disaster detection;

[0187] Database layer: stores all data content in the real-time service system for flood disaster information, among which the SOS database stores the flood sensor model and its monitoring data content; the Feeder database stores the linked SOS data source and the sensor information that needs to be connected to SES; the SES subscription management library manages event subscription information in the form of files; the flood database stores all object information involved in the business logic service layer of this system.

[0188] 3. System development framework

[0189] Adopting the MVC development model, the system is divided into the Model layer, the View layer and the Controller layer. The View represents the front-end interface of the system, while the Model and Controller are responsible for the implementation of the back-end business logic of the system. The front-end interface and the back-end service are separated from each other, and the two are connected through the JSON data interface.

[0190] The front-end interface is developed using Ant-Design components, and virtual DOM components are built based on the React framework. It provides web page operation and display controls such as directory, layout, basic controls, and animation. The front-end framework runs independently with the support of Nodejs. When the web server sends the routing addressing request to the back-end server, the back-end server parses and filters the route and forwards it to the system front-end. First, the mapped page layout is searched through the routing layer, and then the page component layer is loaded. The service JSON interface provided by the back-end service is called, and finally the page information is rendered and displayed to the web server.

[0191] The system backend service is developed using the SpringMVC+Hibernate framework. The backend service Controller layer corresponds to the user interaction layer, providing the system frontend with a JSON data interface that complies with business logic operations; the Service layer corresponds to the business layer in the backend service design, and obtains data information from the object (DataAcccessObject, DAO) layer and WebAPI by calling the database to complete algorithm implementation and business logic processing, thereby providing services to the Controller layer; the DAO layer is directly connected to the database through Hibernate, and the objects obtained from the database by the DAO layer are stored as DO objects in the system unit, and are converted into data transfer objects (DataTransferObject, DTO) after use and transformation by the Service layer, which are used by the Controller layer to display to the frontend;

[0192] (III) System module design

[0193] The system modules are divided into six major modules: user registration / login, flood sensor access unit management, flood sensor map display, flood event subscription, flood event management, and active services during the flood stage. Each module is divided into several small operations, which are independent of each other or connected in sequence, and together constitute the business processing flow of the real-time service system of the large model of flood events.

[0194] 1) User registration / login module: It is based on the "subscribe-publish" service of flood events and provides user registration, login and logout operations to complete specific operation functions of specific users.

[0195] 2) Flood sensor access unit management module: Sensor access is based on topology units. It runs independently on the Storm platform. The management of sensor access units is actually based on the topology unit management based on the StormAPI interface. Through StormAPI, it provides specific users with the functions of viewing, deleting, adding sensor access units, and restarting all.

[0196] 3) Flood sensor map display module: Flood sensors are hydrological and meteorological observation units in geographical scenarios. When displaying sensors to users, the sensor layers are superimposed on the map to intuitively show users the spatial distribution of sensors. By clicking on the sensor icon, users can view basic sensor information, as well as basic map operations such as map zooming in, map zooming out, and personal positioning;

[0197] 4) Flood subscription event module: Based on the flood event subscription model, when users subscribe to flood events, they configure the flood subscription model parameters, including the selection of data sets, the configuration of flood disaster calculation model parameters, and the configuration of flood prediction model parameters. After the subscription configuration parameters are completed, the user registers the flood event to the system by submitting the registration method to complete the flood event subscription;

[0198] 5) Flood event management module: After completing the flood event subscription, users need to manage the flood event content. In this module, users can view the basic information of the flood event, view the flood event status, and delete the flood event. When viewing the flood event, users can view the flood service content at each stage of the event, including the early warning information in the preparation stage, the water level forecast information in the response stage, and the flood calculation information in the recovery stage;

[0199] 6) Flood active service module: The service is automatically triggered by the event stage, which is a flood warning, prediction and event calculation service. It sends the calculation results to the user's subscribed mailbox by email, realizing real-time active service of flood events.

[0200] (IV) Database design

[0201] It is divided into conceptual model design and logical model design. The conceptual model is the process of abstraction and expression of data objects in the system from a cognitive perspective; the logical model is the structured expression of the object relationship in the conceptual model, which is the determination and implementation of the database table structure.

[0202] 1) Conceptual model design: The system involves objects such as users, sensors, flood events, and service content, such as Figure 6This is the ER diagram of the conceptual model of the database of this system. Users manage flood subscription events and sensor access topology units. The status information of the sensor topology unit is managed by StormAPI. Only the identifier needs to be stored in the database object, and the user ID needs to be clarified to determine the management object to which the current topology belongs. The sensor access topology unit object contains the access topology identifier and the user ID; the flood subscription event object is a parameter configuration object that describes the flood disaster AI model. The flood subscription event subscription phase parameters are used as data objects separately. In addition to the basic event attributes such as identifier, name, spatial location, and registration identifier in SES, flood events also include four stage identifiers of diagnosis, preparation, response, and recovery; the identifier corresponds to the flood event subscription phase parameter object, which contains the four-stage filtering parameters of the flood disaster calculation model, including the sensor identifier, monitoring attribute identifier, time window, filtering critical value, and frequency attributes of the flood event stage information; the triggering of flood events is driven by the monitoring data of flood sensors. The system needs to store flood sensors and hydrological and meteorological monitoring attribute objects for data display and calculation. The sensor object includes the sensor identification, name, and spatial location attributes; the hydrological and meteorological monitoring attribute object includes the identification, name, and monitoring data unit attributes of the sensor monitoring attributes; when the flood event detection is completed, the system stores the flood event status record and constructs the flood event process object; the flood event status object is used to store flood messages, and its attributes include the detection event identification, subscription event identification, event type name, and event start and end time content; when a flood message is generated, the system automatically triggers the flood event service, and the service content includes early warning service, water level prediction service, and flood calculation service. In order to record the service content and results of flood events, it is necessary to store various service objects, and the record attributes of services in different flood stages are different due to different service contents; the flood warning service record object includes record identifier, subscription event identifier, warning service time and warning information attribute; the water level prediction service record object includes record identifier, subscription event identifier, prediction service time, prediction service time series, prediction water level result series and actual exploration water level result attribute; the flood calculation service generates a flood event process object according to the flood message, and calculates the flood event process object information. The flood calculation service record includes record identifier, subscription event identifier, calculation service time, maximum water level during the flood and its time, flood start time, flood end time, flood event diagnosis, preparation, response and recovery phase start time and recovery phase end time.

[0203] 2) Table structure design: In the PostgreSQL database environment, the logical structure of the database is displayed with table structure design.

[0204] 4. Implementation and use cases of the real-time service system for the large-scale flood event model

[0205] (I) System development environment

[0206] It is implemented in the browser / server mode and designed and developed in a front-end and back-end separation manner. The front end of the system uses AntDesign components developed with the React framework, and the back end of the system uses the Java-based framework SpringMVC+Hibernate. It relies on dependency inversion IoC and annotation classes in the Spring framework to interpret the MVC hierarchical structure. Hibernate is responsible for connecting to the PostgreSQL database for querying and writing flood data information. It uses the Java persistence layer API for database metadata mapping, and directly associates system entity objects with database objects. At the same time, it uses a variety of front-end UI and service components. In addition to using AntDesign components to implement the basic design of the system, the front end also uses React-amap map components for map implementation, and React-Highcharts chart components to enrich data display. The back-end service relies on Storm, sensor monitoring services, sensor event detection services, and connection services to support the operation of the real-time service system of the large model of flood events.

[0207] (II) Experimental results analysis and discussion

[0208] 1. Analysis of flood event investigation results

[0209] In terms of the degree of exploration refinement, the flood disaster information in this application is more refined and the exploration results are better.

[0210] 2. Comparative analysis of water level forecasts for flood events

[0211] In flood water level prediction, this application uses Matlab for comparative experiments, adopts a "rolling" water level prediction method, uses the data points of the previous 100 days as training data, and predicts the water level monitoring data of the next 2 days. Other parameters are the same as the above-mentioned flood disaster information prediction parameters. This application compares the results of using only precipitation prediction and adding surface temperature, evapotranspiration, relative humidity, wind speed, air pressure, temperature, and sunshine. The overall data set of 20** is used for prediction at intervals of 2 days, and a total of 238 groups of results are predicted. The root mean square error RMSE of the prediction results of the two is 0.8803 for the overall root mean square error of the prediction results of the upstream water level station and precipitation data, 2.2081 for the first flood preparation, response and recovery stage, and 0.8045 for the three stages of the second flood; and the root mean square error of the prediction results is 0.6766 when evapotranspiration and relative humidity data are added to the prediction model for training. The root mean square error of the prediction results is 1.0716 during the first flood period, and 0.6853 during the second flood period. Therefore, the introduction of evapotranspiration, surface temperature, and relative humidity data can significantly improve the prediction results and reduce the error of the results in water level prediction.

[0212] 3. Real-time analysis of system active services

[0213] This feature is specifically reflected in two aspects. The first is the time delay in converting data streams into message streams. Since the basic flood data is regularly pulled and injected into the sensor event service to detect data, there is a certain data delay, so the average time for converting data streams into messages is an important content that characterizes the real-time nature of messages. The second is the delay of active services, which refers to the time it takes to convert real-time message content into specific warning, prediction and calculation services after message change detection, that is, the response time of each service unit.

[0214] like Figure 7 As shown in the figure, in this experiment, the message flow and message records triggered by the inserted 366 data were analyzed. The average time for the data flow to be converted into the message flow was 24.553 seconds. After analysis, in 8 services, it took an average of 92.533 seconds to convert from message changes to corresponding service responses, and the service response could be achieved within two minutes. In addition, the process of constructing emails for the warning service took an average of 0.795 seconds; the prediction service took 42.158 seconds; and the calculation service took an average of 1.219 seconds.

[0215] From the perspective of timeliness, within the system, the average time for flood data streams to be converted into flood messages can be controlled within 0.5 minutes, and the service response can be completed within 2 minutes. Among the three services, the prediction service takes the longest time, with an average of 42.158 seconds, and the water level prediction process can be completed within one minute. From the perspective of the entire service process, after the flood data is accessed by the service, if a flood event occurs, the time to obtain the service result is about 159.264 seconds. Therefore, this system can realize the exploration, early warning and water level prediction services of flood event data within 2.5 minutes, which has strong timeliness.

[0216] 4. System service scalability analysis

[0217] This system framework uses middle-tier services, and the middle service layer is supported by the SWE framework standard. It shields low-level details through the middle layer to ensure the scalability of each independent service. At the same time, the SWE framework performs outstandingly in unified storage and processing of multi-source heterogeneous data. This system is based on SWE, which has certain advantages in data type expansion.

[0218] Secondly, the active service of this system adopts the "producer-consumer" model. The content of the system's active service is determined by the content of the flood status message produced by the sensor event service in the middle platform service. Therefore, according to the different stages of the flood disaster, this system can provide a variety of flood active services. To ensure the scalability of the system, this system provides the IPhaseService service interface. Users implement the executorService(StringeventID, DateeventTime, StringeventType) method in this interface, and after configuring the stage service configuration file reserved in the system, they can customize the stage service content.

[0219] In general, based on the flood detection results, this system can detect flood events 1 day or 4 days in advance, and has a good early warning effect. At the same time, the exploration results divide flood disasters into diagnosis, preparation, response and recovery stages, which characterize flood events in a more refined manner and provide a detailed basis for the subsequent calculation and analysis of flood events. From the perspective of flood event water level prediction results, this system adds evapotranspiration, surface temperature, air temperature and other data content on the basis of the original simple use of precipitation data to predict water levels, and uses the BP neural network to "roll" predict the target water level station data for the next two days. The root mean square error of the prediction is 0.6766 overall, which is significantly improved compared to the case of using only precipitation data. From the perspective of the real-time nature of the system's active services, this system has strong real-time characteristics. It can complete the flood status exploration, early warning and prediction services in the Hanchuan area within 2.5 minutes. It has a strong system foundation for rapid response to flood events such as pre-disaster preparation and post-disaster analysis. From the perspective of system scalability, the system's middle-office services and active service modules have a standard basis, so the data type and low-level scalability are good. At the same time, users can modify the service methods or expand the active service content independently based on the current system according to the SWE framework specification and the interface specification of this system.

Claims

1. The AI ​​large model real-time management system for flood disasters in water conservancy sensor networks is characterized by: Construct a real-time service system for a large-scale flood event model based on the SWE framework. From data access to flood sensors to the detection and processing of flood events, based on the OGC SWE sensor network integration framework, establish multi-source heterogeneous flood sensor registration and real-time access and storage of monitoring data, and explore the diagnosis, preparation, response and recovery stages of flood events under unified standards, and issue flood warnings, hydrological forecasts and flood event calculation services in real time; 1) Establish a large AI model for flood disasters: Based on hydrological and meteorological monitoring data, flood events are divided into four stages: diagnosis, preparation, response, and recovery. The stage status of flood events is calibrated, and the process stages of flood events are refined to provide a basis for early warning and high-granularity services for flood events. In addition, the establishment of a flood event process model includes two aspects: calculation and prediction. The calculation module explores flood information from historical and current hydrological monitoring data, and the prediction module predicts future flood change trends based on historical and current flood monitoring data. 2) Establish a real-time service framework for a large model of flood events based on the SWE framework: Under the SWE standard, realize the automatic access of flood sensor monitoring data, automatic detection of flood events, and automatic triggering of flood stage services. Services are connected to each other based on data interfaces. The framework middle platform actively provides users with flood event detection results and related service content in the form of independent service operation. Flood warning, water level prediction and flood event calculation services are provided for the message content of the preparation, response and recovery stages of flood events. Users obtain flood status information actively pushed by the system through emails or network notifications. In terms of the scalability of multi-source data, two heterogeneous monitoring station data of meteorological stations and hydrological stations are connected. The system middle platform service ensures that the services in the framework work in coordination with the data interface, expands the low-level content of each service, and based on the active service mechanism of messages, ensures that the system can provide corresponding services according to the message status. Users expand different service contents according to the given interface specifications; 3) Design of real-time service system for large-scale flood event model: Using flood data streaming access unit, based on Storm platform, real-time access to flood sensor monitoring data, and real-time release of sensor monitoring service, flood sensor monitoring data flows through SOSSESFeeder link service to sensor event server in real time, real-time detection of flood event status, based on real-time access to monitoring data, real-time detection of flood events, and real-time release of flood status through network notification service, triggering real-time active service after flood status change detection; 4) Based on the BP neural network to predict water level, new meteorological factors such as temperature, air pressure, and atmospheric humidity are added to the prediction method based on precipitation and upstream water level, and the prediction accuracy should be improved.

2. According to claim 1, the water conservancy sensor network flood disaster AI large model real-time management system is characterized in that: Flood disaster AI big model: The model is a rule for flood event detection. Based on the direct characterization factors of the current flood, including water level, flow and other information, the factor characteristics of each stage of the flood event are defined, and the flood event information process model is defined based on the characteristic pattern of water level changes. The flood disaster AI big model consists of two parts. The first part is a characterization model that calculates whether the monitoring frequency of the water level monitoring value in a given interval is greater than or equal to the critical value from the past to the current time range. As a flood disaster calculation model, based on the water level as a direct characterization factor of floods, flood events are divided into four stages: diagnosis, preparation, response and recovery, expressing the flood status from now to the past TW time range; the second part is a flood disaster prediction model that predicts the water level. Based on historical water level and meteorological data, under the premise of factor normalization, the prediction model is trained based on the BP neural network to achieve water level prediction. Based on the flood disaster calculation model in the first part, the future water level prediction function is provided when the flood disaster is in the preparation response and recovery stages. Based on the water level prediction, the future flood trend is analyzed; Define W to represent water level data, U(N) to represent the critical water level interval of stage N, where N∈[A, B, C, D], A represents the diagnosis stage, B represents the preparation stage, C represents the response stage, and D represents the recovery stage. TW(N) represents the time window of stage N, and Fre(TW(N)) refers to the event frequency of the water level WEU(N) of stage N within the TW(N) time window. F(N) represents the critical value of the calculated frequency of stage N. The algorithm steps are as follows: Step 1: Input water level monitoring data; Step 2: When judging the water level W∈U(N), the value space P of N; Step 3: Within the value range of P, calculate the frequency Fre(N) of the event type in the TW(N) time window and compare whether it is greater than F(N), where N∈P. If the calculated frequency is greater than the given value, record the object and further obtain the value space E of N. Step 4: loop through each value in the value space E of N; if the value is A, output it directly; if the value is B, determine whether the state of the previous flood event is A, if so, output stage B, otherwise discard; if the value is C, determine whether the state of the previous flood event is B, if so, output stage C, otherwise discard; if the value is D, determine whether the state of the previous flood event is C, if so, output D, otherwise discard; Step 5: According to the flood stage output in step 4, determine whether the flood event is in the preparation, response and recovery stages; if so, proceed to step 6 to provide flood disaster prediction services; Step 6: Select the historical water level and meteorological data within the TW1 time range, normalize them to form training data, and train the BP neural network model; construct prediction data based on the current monitoring data within the TW2 time range, predict the water level, and obtain the prediction result R.

3. According to claim 1, the water conservancy sensor network flood disaster AI large model real-time management system is characterized in that: Flood disaster calculation model: It is divided into diagnosis, preparation, response and recovery stage filtering conditions. The exploration of each stage must meet two basic filtering conditions: hydrological monitoring data filtering and event stage dependency filtering; Among them, each flood event stage needs to meet the basic hydrological monitoring data filtering conditions: Condi=water level∈U&Fre(TW)≥F Formula 1 The frequency Fre of the water level monitoring data shown in Formula 1 appearing in a given water level interval U within a certain time TW is greater than or equal to the given frequency F; Formula 2 is the dependent filtering condition of the flood event stage, which is the key to ensure the mutual dependence of each stage of the flood event. The definition of the flood event stage must strictly meet the sequential dependence relationship of A->B->C->D. The definition of the current stage CurrentPhase must be based on meeting the filtering conditions of the hydrological monitoring data and strictly depend on the previous exploration results; CurrentPhase=Condi after LastPhase, LastPhase∈[A, B, C] Equation 2 The filter condition A that the monitoring data in the diagnosis phase should meet is: A=water level<W1&Fre(TW=T1)≥F1 Formula 3 The filter condition B that the monitoring data in the preparation phase should meet is: BCondi=W1≤water level <W2&Fre(TW=T2)≥F2 B=Bcondi afterA Formula 4 The filtering condition C that the monitoring data in the response phase should meet is: Ccondi=water level≥W2&Fre(TW=T3)≥F3 C=Ccondi after B Formula 5 The filtering condition D that the monitoring data in the recovery phase should meet is: Dcondi=water level<W3&Fre(TW=T4)≥F4 D=Ccondi after C Formula 6 W1, W2 and W3 are water level values ​​at different stages, and W1<W2 and W3<W2 are satisfied; TW stands for TimeWindow, which refers to the time window. T1, T2, T3 and T4 are different critical values ​​of the time window, and T1≥T2≥T3≥T4 is satisfied; Fre(TW) refers to the number of times the qualified water level value appears in the time window TW, and F1, F2, F3 and F4 are different number values. The relative size restrictions of the water level value and the length of the time window are used to avoid confusion between different stages of the flood event and ensure the uniqueness of the stage of the flood event; in addition, the diagnosis, preparation, response and recovery stages are strictly in sequence. In addition to the need to judge whether the current basic conditions Bcondi, Ccondi and DCondi are met, the three conditions B, C and D also need to meet the change trend of the evolution of the A->B->C->D stages, that is, the preparation stage must be after the diagnosis stage, the response stage must be after the preparation stage, and the recovery stage must be after the response stage.

4. According to claim 1, the water conservancy sensor network flood disaster AI large model real-time management system is characterized in that: Flood disaster prediction model: The computational model characterizes the changing characteristics of hydrological monitoring data from the past to the present, and the prediction model performs hydrological prediction of flood events based on the flood event status information extracted after preliminary filtering by the computational model, thereby improving the accuracy and timeliness of the re-judgment of the occurrence time and status of flood events; The short-term water level prediction model is constructed based on BP neural network, which consists of two parts: forward feedback and backward feedback. In the forward feedback process, the input vector X passes through the input layer, hidden layer, and output layer to output vector Y; The backward feedback process outputs the vector Y according to the forward feedback, and the target vector Y target The error content between them is used to reversely solve the weight adjustment gradient of the output layer and the hidden layer to form a new numerical model. The specific algorithm flow is as follows: Process 1: Build the network structure: Assume that there are n nodes in the input layer (the number of nodes in the input layer is equal to the length of the input vector), p nodes in the hidden layer, and q nodes in the output layer (the number of nodes in the output layer is equal to the length of the output vector); Process 2: Let the input vector be X = [x1, x2, ..., x n ], hidden layer vector H = [h1, h2, ..., h p ], the output layer vector Y = [y1, y2, ..., y q ], the expected output vector T = [t1, t2, ..., t q ], the weight between the input layer and the hidden layer is W I , B I is the bias vector between the two layers: the weight between the hidden layer and the output layer is W 2 , B 2 is the bias vector between the two layers; Process 3, after initializing the weight W and the bias vector B, the expression for calculating the hidden layer vector from the input layer vector is: H=f(W 1 X+B 1 ) Formula 7 The expression for calculating the output layer vector from the hidden layer vector is: Y=f(w 2 H+B 2 ) Formula 8 Thus, the calculation of the output vector Y of the forward feedback is completed; Process 4: Calculate the error function between the expected output vector T and the vector Y, and reversely derive the weight adjustment gradient of the output layer and the hidden layer based on the relationship between the input layer, hidden layer, and output layer described in process 3. The error function is: The error is back-propagated from the output layer to the hidden layer, and the weight adjustment of each layer is updated using the gradient descent method, that is, W = W-ηΔE p (W, B), where η is the set learning rate, is the step size of each update, and ΔE p (W, B) is the partial derivative of the output deviation under the p-th sample input with respect to the weight of a certain layer, indicating that the parameter is updated once each sample is input; Process 5: Based on the above weight adjustment method, calculate the new weight value, implement the backward feedback process of the neural network, and update the weight value; Process 6: Calculate the global error and determine whether the network error meets the requirements. When the error reaches the preset accuracy or the number of learning times is greater than the set maximum number, the algorithm ends. Otherwise, select the next learning sample and the corresponding expected output, return to process 3, and enter the next round of learning. When constructing a flood prediction model, this application adds factors such as evapotranspiration, air humidity, temperature, and air pressure as inputs based on the BP neural network model to improve the accuracy of water level prediction; The flood disaster prediction model is a service module for re-judging flood events based on the flood disaster calculation model. This part provides prediction services during flood disasters based on the flood event preparation and response stage status detected by the flood disaster calculation model, calculates the water level data for the next two days, and then judges the subsequent flood trend based on the stage water level critical value in the flood disaster calculation model. On this basis, the BP neural network water level prediction service is realized to improve the accuracy of stage detection of flood events.

5. According to claim 1, the water conservancy sensor network flood disaster AI large model real-time management system is characterized in that: Real-time service framework for large-scale flood event models: It is a highly timely flood service framework that covers everything from real-time acquisition and release of flood data to real-time exploration and calculation of flood status, and then to detection of flood status changes and active service response. The framework is based on sensor monitoring services and event services, constructs sensor access methods and flood active service models, and completes real-time flood disaster information service methods. The framework consists of four parts: flood data basic services, flood subscription service registration center, flood disaster detection, and flood active services. The flood data basic service is the driving service of the framework. The sensor streaming data access unit accesses the sensor monitoring data in real time, converts the sensor byte stream into the data stream under the SWE framework, and inserts it into the sensor monitoring service in real time. The sensor monitoring service exposes the accessed flood data records in real time through the interface and publishes the data content in real time. The flood subscription service registration center is the flood event subscription interface provided by the framework. Users can define the flood disaster AI model through the subscription service center. The flood disaster calculation model defined by the user in the flood subscription service registration center is encoded in EML and acts on the sensor event service to provide a detection model for flood message production. The defined flood disaster prediction service acts on the flood active service, clarifies the water level prediction model structure, and assists in the re-judgment of future flood disaster information. The sensor event service is the core of flood disaster detection. The event server filters the specific monitoring data pulled from the sensor monitoring service by the link layer according to the flood mode subscribed by the registration center, detects the flood disaster status information, publishes the message object, and forms a message queue. The message object Message includes the flood event identifier EventID, the flood event detection time DetectedTime, and the flood event status type EventPhaseType; Message=[EventID, DetectedTime, EventPhaseType] Equation 10 The flood active service automatically converts the flood stage status into flood information service content based on the flood message content generated by flood stage exploration, and is responsible for consuming flood messages. In the active service module, the message queue is detected through changes, and the flood active service type and content are triggered and changed in real time. When the diagnosis is transformed into the preparation stage or the preparation is transformed into the response stage, the scheduled flood forecast and warning service is triggered; when the response is transformed into the recovery stage, the flood status degradation information service is released; when the recovery is transformed into the diagnosis stage, that is, at the end of the recovery stage, flood disaster information calculation services are provided.

6. According to claim 1, the water conservancy sensor network flood disaster AI large model real-time management system is characterized in that: Sensor access method with storage and computing separation: data reception, data filtering and data storage are separated from each other. Each component unit is independent of each other and has high reusability. Based on the Storm streaming processing framework, sensor data access is divided into data reception component, monitoring filtering component and monitoring storage component. The access framework takes the Storm topology monitoring access unit in the topology layer as the core, converts the byte stream data of the sensor network layer into a data stream that meets the OGCSWE monitoring service standard. With the support of the SOS service, the sensor monitoring data can be quickly acquired, parsed, filtered and stored. 1) Data receiving component: Under the OGC SWE standard, the data receiving component obtains the sensor eight-tuple metamodel from the SOS service, builds a unified access model in a sensor site-centric manner, and organizes information such as the sensor's geographic spatial location, monitoring time, and monitoring attributes under the sensor site. The unified access model is an information storage unit for sensor attribute monitoring values, and describes sensor monitoring in the form of "sensor site-sensor-monitoring attribute". Sites and sensors, sensors and monitoring attributes are organized in a one-to-many manner. In addition to building a unified sensor access model, the data acquisition component obtains and parses the byte stream transmitted by the sensor from the network layer, and obtains the corresponding sensor data transmission protocol and data expression-related data access metadata from the sensor metamodeling. The structure of the data expression is in the form of an array of [SensorID, PropertyID, DataPosition, DataLength, DataRatio], where SensorID refers to the current sensor ID, PropertyID is the sensor monitoring attribute ID, DataPosition is the data position, DataLength is the data length, and DataRatio is the ratio of the monitoring data converted to the target unit; 2) Monitoring and filtering components: Sensor monitoring is an expression of geographical phenomena related to space and time. The monitoring and filtering components solve the monitoring scene limitations encountered during the sensor monitoring access process, and filter the time, space and monitoring values. The filtering module supports multi-condition filtering based on the OGC Filter standard. It connects each filtering unit with AND logic operations, and supports monitoring value filtering in spatiotemporal scenarios. Spatial filtering is filtering for sensor spatial scenes, and filtering for spatial points, including methods such as inclusion, non-inclusion and rectangular query. Time filtering is filtering for monitoring time scenes, filtering time point information, including operations within a certain time range, before a certain time and after a certain time. Attribute filtering is filtering for abnormal values, and supports basic comparison filtering operations such as greater than, less than and equal to. 3) Monitoring storage component: After the sensor monitoring data is collected, parsed, and filtered, its data format and expression meet the requirements of the corresponding application scenario. At this time, the monitoring storage component converts its data structure into a standard monitoring data model structure encoded in O&M, and constructs a monitoring data service insertion request to realize the rapid insertion process of sensor monitoring data; The sensor data receiving component, monitoring filtering component and monitoring storage component are the basic units for sensor monitoring data access. The Storm streaming data processing framework integrates the three components on this basis, constructing the data receiving component with the Spout unit, and the monitoring filtering component and monitoring storage component with the Blot unit, forming a sensor monitoring access unit that meets the Storm platform standard. The topology standardizes the sensor access process and ensures that the sensor access has the possibility of real-time change. The access process uses the topology template to build the topology in the XMLSchema standard. The configuration mode includes four parts: topology identifier TopologyID, sensor access component Receive, sensor filtering component Filter and sensor monitoring storage component Store. After the topology node is uploaded and started, the sensor access unit enters the running state. The state management is controlled by the RestflAPI service provided by StormUI to turn on and off the sensor access topology unit.

7. According to claim 1, the water conservancy sensor network flood disaster AI large model real-time management system is characterized in that: Active flood service: provides three active services: early warning, water level prediction and flood calculation. Flood early warning, alarm and downgrade release service is a continuous phase service acting on the active service module, corresponding to the water level prediction service in the active flood service; the flood water level prediction service starts from the flood preparation stage and ends automatically after the flood recovery stage. In the preparation stage, the system provides flood early warning and prediction services; in the response stage, it also provides flood early warning and flood prediction services; Providing flood downgrade notifications, flood forecasts, and flood calculation services during the recovery phase; 1) Flood warning service: Publish flood status information in real time in the form of emails. The system automatically edits flood event warning information content based on the current flood status, calls the email service through the email server, notifies users of flood event status information, and generates flood warning records; 2) Flood prediction service: Based on the current water level of the upstream water level station and the meteorological data information in the basin, the water level data of the target monitoring station in the next two days is predicted in real time. The steps are as follows: Step 1: Enter the current event ID and time, obtain the parameters of the flood disaster AI model, and obtain the prediction model parameter configuration; Step 2: According to the prediction model parameter configuration, obtain flood monitoring data and align the data content to form prediction model training data and prediction input data; Step 3: Normalize the training data, initialize the BP neural network model parameters, and train the neural network model; Step 4: Normalize the predicted data and use the trained neural network model to predict the water level to obtain the normalized prediction results; Step 5: Based on the normalized prediction results, perform denormalization to obtain the water level prediction results and store the prediction results.

8. According to claim 1, the water conservancy sensor network flood disaster AI large model real-time management system is characterized in that: Flood calculation service: Flood disaster objects refer to the complete flood event objects detected in the flood disaster information service, including the start and end time of the flood event, the start time of the flood diagnosis phase, the start time of the preparation phase, the start time of the response phase, and the start and end time of the recovery phase, describing the information of each stage of the entire flood event process. The flood calculation service calculates flood disaster information, automatically generates, stores, and manages flood event process objects, and calculates the maximum water level information during the flood occurrence stage, so as to prepare data for subsequent flood event analysis and research; The flood disaster calculation service process steps are as follows: Step a: Based on the flood event ID and the current time point, obtain the historical message time series set from the end of the previous event to the current event, and initialize the flood disaster object; Step b: According to the message type classification, the current message is divided into four stages: diagnosis, preparation, response and recovery; Step c: If the current message type is the diagnosis stage, determine whether the message type has changed. If it has changed, update the flood event status, otherwise update the preparation stage and event start time of the flood disaster object; and determine whether the previous message type is the recovery stage. If so, set the flood disaster object recovery stage and event end time to store the flood disaster object and calculate the maximum water level. Otherwise, update the preparation stage and event start time of the flood disaster object; if the current message type is the preparation, response or recovery stage, update the current event status and set the start time of the corresponding stage, check whether the message set is traversed, if so, throw an exception, otherwise, return to step b.

9. According to claim 1, the water conservancy sensor network flood disaster AI large model real-time management system is characterized in that: Business process of the real-time service system for large-scale flood event models: including four business parts: user subscription to flood events, large-scale flood event model exploration and service, sensor access management, and flood event management and review; In the user subscription service, users configure the flood event subscription model, select the monitoring data set related to the flood event from the sensor monitoring service, and then configure the flood disaster calculation model parameters, flood disaster prediction model and other parameters. After that, the flood event subscription model is registered in the sensor event service and managed uniformly in the flood database server; In the business of large-scale model exploration and service of flood events, hydrological and meteorological sensors, under the premise of modeling registration, access sensor monitoring data in real time through sensor access units, and publish monitoring data in real time through sensor monitoring services. After the monitoring data is uniformly released by SOS, it flows into the sensor event exploration server, explores each flood event stage in real time, triggers the service modules of each stage of flood events, and actively sends service results to users in the form of emails, completing the active exploration and service of flood events. The exploration results and stage service results will also be recorded in the flood database to serve the subsequent flood management and information display business; In the sensor access management service, users can view, delete, and add sensor access units to achieve visual management of access sensors. In the flood management and viewing service, users can obtain flood history content through sensor monitoring services, sensor event services, and flood databases, and actively view flood sensor information, flood sensor monitoring data information, flood event subscription information, flood event stage status information, and flood stage service information. Based on the above business processes, the system carries out overall framework design, functional design and database design, and clarifies the main functions, task processes and data structures of the real-time service of the large model of flood events.

10. The water conservancy sensor network flood disaster AI large model real-time management system according to claim 1 is characterized in that: System backend service: adopts a four-layer hierarchical system, dividing the backend into user interaction layer, business logic layer, intermediate service layer and database layer; User interaction layer: provides corresponding interface support for the operation units in the system interface, and directly relies on the business logic layer of the lower layer, that is, the business logic layer provides specific algorithm implementation for the user interaction layer, and the front-end interface calls the user interaction layer interface to complete the business process; the operation interface of this layer includes login interface, map operation interface, flood management interface, flood sensor access unit management interface and flood subscription model parameter configuration interface, and the active service interface directly sends emails to users through email services to notify the stage status of flood events and water level prediction results; with the support of the interface, users can log in to this system and upload sensor access units by themselves to obtain data in specific scenarios; at the same time, users subscribe to flood events in a certain geographical scenario and view the corresponding flood disaster information through the flood event management unit; Business logic layer: responsible for calling the service content of the middle service layer, and after parsing or encoding the data, integrating the business process according to business needs, and providing the required data objects and algorithm results to the user interaction layer and the middle service layer. The business logic includes flood event subscription coding, OGCSWE standard parsing, login logic judgment, sensor monitoring access and flood stage services; Middle service layer: abstracts the basic service units in the system into SOA services, bringing business reuse. At the same time, each service unit in the middle service layer can expand its performance. Sensor monitoring uses stand-alone services based on PostgreSQL storage or distributed services based on MongoDB storage. According to the principle of active service module, it can actively detect flood events and trigger flood service content in the corresponding stage; the sensor access unit runs in real time on the Storm platform, registers the flood data content perceived by the flood sensor in real time to the sensor monitoring service, and the Feeder link service regularly pulls the flood data content required by the sensor event detection service and injects it into the sensor event server in real time. The sensor event server can actively detect flood events subscribed by users, and finally actively publish flood event status information through the network notification service to complete the automated flood disaster detection; Database layer: stores all data content in the real-time service system for flood disaster information, among which the SOS database stores the flood sensor model and its monitoring data content; the Feeder database stores the linked SOS data source and the sensor information that needs to be connected to SES; the SES subscription management library manages event subscription information in the form of files; the flood database stores all object information involved in the business logic service layer of this system.

Images