A method and system for constructing a water conservancy big data service analysis and evaluation model

Abstract

The present application relates to the field of model building and analysis technology, and discloses a method and system for building a water conservancy big data service analysis and evaluation model. The method includes: extracting feature vectors from water conservancy monitoring data through a dual-stream attention mechanism, constructing a dynamic heterogeneous graph neural network based on the feature vector to generate a system graph representation, using the feature vector and graph representation to perform multi-task prediction to obtain prediction results, optimizing the prediction results through hierarchical reinforcement learning to generate a scheduling plan, performing fuzzy neural network analysis on the scheduling plan to obtain an evaluation result, and finally updating the optimized service model through data stream processing. The present application realizes the dynamic construction and continuous optimization of the water conservancy big data service analysis and evaluation model.

Description

Technical Field

[0001] The present application relates to the field of model building and analysis technology, and in particular to a method and system for building a water conservancy big data service analysis and evaluation model. Background Art

[0002] In the field of water conservancy big data analysis, traditional service analysis and evaluation models are mainly built based on expert experience and fixed rules. These models usually use methods such as hierarchical analysis and fuzzy comprehensive evaluation to evaluate water conservancy scheduling plans, and quantify the evaluation results by setting a fixed evaluation index system and weight coefficients. At the same time, machine learning methods such as support vector machines and artificial neural networks have also begun to be introduced into existing technologies to deal with the prediction and classification of hydrological data, and combined with GIS technology to realize the visualization and spatial analysis of water conservancy information.

[0003] However, the existing technology has the following shortcomings: the fixed evaluation index system is difficult to adapt to the dynamic changes of the water conservancy system and cannot capture the emerging data patterns and business needs in a timely manner; the single machine learning model has limited generalization ability when processing multi-source heterogeneous water conservancy data, and it is difficult to take into account both temporal and spatial characteristics at the same time; the traditional batch learning method has low computational efficiency and is not suitable for processing the continuously growing water conservancy big data; once the evaluation model is established, it is difficult to update it dynamically and cannot be adaptively adjusted with changes in the environment and business needs; in the process of multi-objective scheduling optimization, it is difficult to balance the needs of different stakeholders, resulting in limited practicality of the scheduling plan. Summary of the invention

[0004] The present application provides a method and system for constructing a water conservancy big data service analysis and evaluation model, which is used to realize the dynamic construction and continuous optimization of the water conservancy big data service analysis and evaluation model.

[0005] In the first aspect, the present application provides a method for constructing a water conservancy big data service analysis and evaluation model, which includes: performing multimodal feature extraction on water conservancy monitoring data through a dual-stream attention mechanism to obtain a water conservancy data feature vector; constructing a dynamic heterogeneous graph neural network based on the water conservancy data feature vector, and generating a water conservancy system graph representation through an adaptive graph convolution operation; using the water conservancy data feature vector and the water conservancy system graph representation, performing multi-task prediction on water conservancy indicators through a shared-private feature extractor to obtain a water conservancy prediction result; based on the water conservancy prediction result, optimizing and calculating water conservancy scheduling through a hierarchical reinforcement learning framework to generate a water conservancy scheduling plan; based on the water conservancy scheduling plan, performing a comprehensive analysis of the water conservancy system through a fuzzy neural network to obtain a water conservancy evaluation result; based on the water conservancy evaluation result, performing incremental updates through a data stream processing engine to obtain an optimized water conservancy service model.

[0006] In a second aspect, the present application provides a water conservancy big data service analysis and evaluation model construction system, the water conservancy big data service analysis and evaluation model construction system comprising:

[0007] The extraction module is used to extract multimodal features of water conservancy monitoring data through a two-stream attention mechanism to obtain a water conservancy data feature vector;

[0008] A generation module, used for constructing a dynamic heterogeneous graph neural network according to the water conservancy data feature vector, and generating a water conservancy system graph representation through an adaptive graph convolution operation;

[0009] A prediction module, used to use the water conservancy data feature vector and the water conservancy system diagram representation to perform multi-task prediction on the water conservancy index through a shared-private feature extractor to obtain a water conservancy prediction result;

[0010] A calculation module, used to optimize the water conservancy dispatching through a hierarchical reinforcement learning framework based on the water conservancy prediction results to generate a water conservancy dispatching plan;

[0011] An analysis module is used to conduct a comprehensive analysis of the water conservancy system through a fuzzy neural network according to the water conservancy dispatching plan to obtain a water conservancy evaluation result;

[0012] The updating module is used to perform incremental updates through a data stream processing engine according to the water conservancy evaluation results to obtain an optimized water conservancy service model.

[0013] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned water conservancy big data service analysis and evaluation model construction method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the above-mentioned water conservancy big data service analysis and evaluation model construction method.

[0015] In the technical solution provided by this application, multimodal feature extraction is performed on water conservancy monitoring data through a dual-stream attention mechanism, which effectively solves the problem of unified processing of structured data and unstructured data, and realizes the deep fusion of temporal features and spatial features. Based on the extracted feature vectors of water conservancy data, a dynamic heterogeneous graph neural network is used to construct an overall representation of the water conservancy system. The dynamic correlation between water conservancy facilities is effectively captured through adaptive graph convolution operations, and the expression ability of the structural characteristics of complex water conservancy systems is enhanced. The introduction of a shared-private feature extractor for multi-task prediction not only ensures the sharing and transfer of knowledge between different prediction tasks, but also maintains the specific characteristics of each task, thereby improving the prediction accuracy. A hierarchical reinforcement learning framework is used for water conservancy scheduling optimization, and complex scheduling problems are decomposed into multiple subtasks. Global optimal scheduling is achieved through collaborative learning of intelligent agents. The application of fuzzy neural networks enables the evaluation process to simultaneously process qualitative and quantitative indicators, improving the scientificity and interpretability of the evaluation results. The data stream processing engine realizes online learning and continuous optimization of the model, and enables the model to adapt to the dynamically changing water conservancy environment through an incremental update mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0017] Figure 1 This is a schematic diagram of an embodiment of a method for constructing a water conservancy big data service analysis and evaluation model in an embodiment of the present application;

[0018] Figure 2 A schematic diagram of a process for performing layered decomposition in an embodiment of the present application;

[0019] Figure 3 A schematic diagram of an evaluation factor set in an embodiment of the present application;

[0020] Figure 4 A schematic diagram of an embodiment of a water conservancy big data service analysis and evaluation model construction system in an embodiment of the present application;

[0021] Figure 5 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The embodiment of the present application provides a method and system for constructing a water conservancy big data service analysis and evaluation model. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0023] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the method for constructing a water conservancy big data service analysis and evaluation model in the embodiment of the present application includes:

[0024] Step S101: extract multimodal features from water conservancy monitoring data through a dual-stream attention mechanism to obtain a water conservancy data feature vector;

[0025] Step S102: construct a dynamic heterogeneous graph neural network according to the water conservancy data feature vector, and generate a water conservancy system graph representation through an adaptive graph convolution operation;

[0026] Step S103, using the water conservancy data feature vector and the water conservancy system diagram representation, a shared-private feature extractor is used to perform multi-task prediction on the water conservancy index to obtain a water conservancy prediction result;

[0027] Step S104: Based on the water conservancy prediction results, the water conservancy dispatch is optimized and calculated through a hierarchical reinforcement learning framework to generate a water conservancy dispatch plan;

[0028] Step S105: Based on the water conservancy dispatching plan, a comprehensive analysis is performed on the water conservancy system through a fuzzy neural network to obtain a water conservancy evaluation result;

[0029] Step S106: Based on the water conservancy evaluation results, incremental updates are performed through the data stream processing engine to obtain an optimized water conservancy service model.

[0030] It is understandable that the execution subject of this application can be a water conservancy big data service analysis and evaluation model construction system, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0031] Specifically, the multi-source heterogeneous water conservancy monitoring data is processed. The dual-stream attention mechanism contains two parallel processing branches: the structured data branch processes numerical data such as water level, flow, and rainfall of hydrological stations and rain gauges, and the unstructured data branch processes image data such as satellite remote sensing images and video surveillance. Each branch diverts the input data through a temporal convolutional network. The temporal attention module captures the dynamic change characteristics at different time scales, and the spatial attention module learns the spatial dependencies between different monitoring points. Taking the monitoring data of the hydrological station as an example, the water level, flow and other data of each monitoring point are feature decomposed according to multiple time scales such as hours, days, and months to obtain a temporal feature matrix reflecting the change laws of different time periods. For remote sensing image data, spatial features of different regions, such as water area, vegetation coverage and other information, are extracted through regional adaptive division. The features of the two branches are fused through cross-attention operations, and positive sample pairs are constructed for comparative learning to obtain the water conservancy data feature vector. After obtaining the water conservancy data feature vector, a dynamic heterogeneous graph neural network is constructed to characterize the overall structure of the water conservancy system. The water conservancy facilities (such as reservoirs, hydropower stations, pumping stations, etc.) are taken as nodes of the graph, and the physical connection relationships between different facilities (such as upstream and downstream relationships, water pipe network connections, etc.) are taken as edges to construct the initial graph structure. The weight of the edge is determined by correlation calculation to reflect the degree of mutual influence between the facilities. The physical relationship between nodes and edges is encoded, and the attribute information (such as reservoir capacity, power generation capacity, etc.) and connection characteristics (such as channel flow, pipe network pressure, etc.) of the water conservancy facilities are embedded into the graph structure. The time-varying graph analyzer captures the evolution characteristics of the graph structure over time through dynamic connection strength calculation, such as the adjustment of scheduling strategies caused by seasonal changes. Construct a multi-level graph representation to extract the hierarchical characteristics of the water conservancy system layer by layer from local to global. Graph matching is performed through topological similarity calculation to ensure the robustness of the graph representation to structural disturbances. Finally, the overall representation of the water conservancy system is generated through adaptive graph convolution operation.

[0032] Multi-task prediction is performed based on the feature vector of water conservancy data and the representation of water conservancy system diagram. Information fusion encoding is performed on the features to generate a fusion matrix containing point features and structural features. The features are decoupled into shared features and private features through symmetric loss constraints and orthogonal decomposition. The shared features reflect the common information of different prediction tasks, and the private features characterize the specific information of each task. The shared features are dynamically grouped based on mutual information calculation, and the tasks with strong correlation are combined together. The weight of each task is determined by feature importance calculation and contribution ranking. A multi-dimensional prediction target space is constructed, and different types of prediction indicators (such as water quantity prediction, water quality prediction, scheduling prediction, etc.) are uniformly modeled. Finally, various prediction results are obtained through feature reorganization and multi-task fusion. For water conservancy scheduling optimization, a hierarchical reinforcement learning framework is adopted. Complex scheduling problems are decomposed into multiple subtasks, such as power generation scheduling, flood control scheduling, water supply scheduling, etc. A state space is constructed for each subtask, including state variables such as water level, flow, demand, and control actions such as gate opening and pump station start and stop. Based on safety constraints (such as flood control limit water level, ecological base flow, etc.), a reward and punishment mechanism is designed to build a reward function that reflects the scheduling target. The idea of ​​predictive control is used for rolling optimization, and the scheduling strategy is dynamically adjusted according to short-term prediction results. Multiple intelligent agents optimize the joint scheduling plan through a collaborative learning mechanism to maximize the overall benefits of the system.

[0033] When comprehensively evaluating the water conservancy system, the water conservancy dispatching plan is hierarchically decomposed according to technical, economic, social, and environmental dimensions to obtain specific evaluation factors. The qualitative and quantitative indicators are uniformly converted into fuzzy features through fuzzy rule mapping, and the membership of each indicator is calculated. The evaluation rule library is constructed in combination with expert knowledge, and correlation analysis is performed on different indicators. The attention mechanism is used to learn the indicator weights to avoid subjective human empowerment. A comprehensive score is obtained through the fusion of multi-level indicators to generate the water conservancy evaluation results.

[0034] Continuous optimization of the model is achieved through data streaming processing. Real-time monitoring and incremental feature analysis of water conservancy evaluation results are carried out to detect dynamic changes in data distribution. Concept drift detection is used to timely detect system performance degradation and generate update trigger signals. Incremental learning sample sets are constructed to evaluate the value of new samples and select high-quality samples for model updates. Knowledge continuity is maintained through progressive training to continuously optimize the performance of water conservancy service models.

[0035] In the embodiment of the present application, multimodal feature extraction is performed on water conservancy monitoring data through a dual-stream attention mechanism, which effectively solves the problem of unified processing of structured data and unstructured data, and realizes the deep fusion of temporal features and spatial features. Based on the extracted water conservancy data feature vector, a dynamic heterogeneous graph neural network is used to construct an overall representation of the water conservancy system. The dynamic correlation between water conservancy facilities is effectively captured through adaptive graph convolution operations, and the expression ability of the structural features of complex water conservancy systems is enhanced. The introduction of a shared-private feature extractor for multi-task prediction not only ensures the sharing and transfer of knowledge between different prediction tasks, but also maintains the specific characteristics of each task, thereby improving the prediction accuracy. A hierarchical reinforcement learning framework is used for water conservancy scheduling optimization, and complex scheduling problems are decomposed into multiple subtasks. Global optimal scheduling is achieved through collaborative learning of intelligent agents. The application of fuzzy neural networks enables the evaluation process to simultaneously process qualitative and quantitative indicators, improving the scientificity and interpretability of the evaluation results. The data stream processing engine realizes online learning and continuous optimization of the model, and enables the model to adapt to the dynamically changing water conservancy environment through an incremental update mechanism.

[0036] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0037] (1) The water conservancy monitoring data is processed through the time series convolutional network to generate structured data streams and unstructured data streams;

[0038] (2) Input the structured data stream into the temporal attention module and obtain the temporal feature matrix through multi-scale feature decomposition;

[0039] (3) Input the unstructured data stream into the spatial attention module and obtain the spatial feature matrix through region adaptive partitioning;

[0040] (4) Perform cross-attention operations on the temporal feature matrix and the spatial feature matrix to obtain a fused feature tensor;

[0041] (5) Construct contrastive learning sample pairs based on the fused feature tensor and obtain positive sample sets through data enhancement;

[0042] (6) Input the positive sample set into the contrast loss function for feature mapping to obtain the water conservancy data feature vector.

[0043] Specifically, the original data is classified through a temporal convolutional network. The network adopts a multi-layer convolution structure and designs specific convolution kernels for the temporal characteristics of water conservancy monitoring data. The size of the convolution kernel gradually increases with the number of layers, thereby capturing temporal patterns of different scales. For structured data such as water level, flow, and water temperature monitored in real time by hydrological stations, their temporal characteristics are retained after convolution processing and a structured data stream is formed; for unstructured data such as satellite remote sensing images and video surveillance, spatial features are extracted through convolution operations to generate unstructured data streams. After the structured data stream enters the temporal attention module, the features are decomposed according to different time scales. Taking the monitoring data of the hydrological station as an example, the hourly data is aggregated into daily averages through a sliding window, and then the daily averages are aggregated into monthly averages to form multi-scale temporal features. The temporal attention calculation can be expressed as:

[0044] ;

[0045] in, represents the temporal attention weight matrix (unit: dimensionless), is the query matrix (reflecting the current moment characteristics), is the key matrix (storing historical moment features), is the value matrix (containing the actual data values), is the feature dimension. This formula is used to calculate the correlation strength between different time points and generate a time series feature matrix. After the unstructured data stream is input into the spatial attention module, the region is adaptively divided. Taking the reservoir remote sensing image as an example, the entire area is divided into different sub-areas such as water area, coastline, vegetation, etc. according to the image features, and the features of each sub-area are extracted through the spatial attention mechanism. The spatial attention calculation is expressed as:

[0046] ;

[0047] in, represents the spatial attention weight matrix (unit: dimensionless), is the region query matrix, is the region key matrix, is the region value matrix, is the spatial feature dimension. This formula is used to calculate the correlation strength between different spatial regions and generate a spatial feature matrix.

[0048] The temporal feature matrix and the spatial feature matrix are fused through cross-attention operation. The fusion process is calculated as follows:

[0049] ;

[0050] in, is the fused feature tensor, and are the weight matrices of temporal and spatial features respectively, and MLP is a multi-layer perceptron, which is used for feature dimension reduction and nonlinear transformation. This fusion operation unifies the temporal and spatial features into the same feature space.

[0051] When constructing contrastive learning samples based on fused feature tensors, data enhancement technology is used to generate positive samples. For hydrological data, similar samples are constructed by adding random noise, time shift, etc.; for remote sensing images, enhanced samples are generated by transformations such as rotation and cropping. These samples and the original data together constitute positive sample pairs for training contrastive learning models. The positive sample pairs are feature mapped through the contrast loss function to obtain the feature vector of water conservancy data. In practical applications, such as real-time monitoring of a reservoir, water level, flow and other data are collected once an hour, and remote sensing images are obtained once a day. After the above processing, a comprehensive feature vector that can simultaneously characterize the hydrological characteristics and spatial characteristics of the reservoir is obtained, providing basic data support for subsequent analysis and prediction.

[0052] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0053] (1) The water conservancy data feature vectors are used to construct a water conservancy node set according to the spatial position relationship, and the water conservancy edge set is obtained by calculating the association degree;

[0054] (2) Encode the physical relationships of the water conservancy node set and the water conservancy edge set to generate the initial graph topology structure;

[0055] (3) Input the initial graph topology into the time-varying graph analyzer and obtain the graph evolution sequence through dynamic connection strength calculation;

[0056] (4) Generate a multi-level graph representation based on the graph evolution sequence, and obtain the water conservancy sub-graph set through hierarchical aggregation operations;

[0057] (5) Perform structural comparison analysis on the water conservancy sub-atlas and obtain the graph matching matrix by calculating the topological similarity;

[0058] (6) The graph matching matrix is ​​input into the adaptive graph convolution operation to generate a water conservancy system graph representation.

[0059] Specifically, the water conservancy data feature vectors are processed and organized into nodes and edges according to the actual spatial position relationship. In the water conservancy system, nodes represent specific facilities such as reservoirs, hydropower stations, pumping stations, etc. Each node contains the characteristic information of the facility, such as the reservoir capacity, water level, inflow, etc. The edge represents the physical connection relationship between facilities, such as upstream and downstream relationships, water pipe network connections, etc. The correlation between nodes is calculated using the following formula:

[0060] ;

[0061] in, represents the degree of association between node i and node j (unit: dimensionless), and Respectively represent the values ​​of the i-th node and the j-th node on the k-th feature dimension, is the weight coefficient of the kth feature, and n is the total number of features. Taking multiple reservoirs in a basin as an example, the correlation between different reservoirs is calculated by this formula, taking into account multiple factors such as water level difference, flow relationship, and spatial distance.

[0062] When encoding the physical relationships of the constructed initial graph structure, the dynamic characteristics of water conservancy facilities need to be considered. The time-varying graph analyzer describes the evolution characteristics of the graph structure over time by calculating the dynamic connection strength. The calculation formula is:

[0063] ;

[0064] in, represents the graph structure state matrix at time t, is the time series smoothing factor (between 0 and 1), is the node feature matrix, is the edge feature matrix, is a nonlinear transformation function. In practical applications, for example, the difference in scheduling strategies between flood season and dry season in a river basin will lead to changes in the connection relationship between water conservancy facilities. This formula can capture this dynamic change feature. Based on the graph evolution sequence, a multi-level graph representation is constructed through hierarchical aggregation operations. The mathematical expression of the aggregation process is:

[0065] ;

[0066] in, is the graph representation of the l+1th layer, represents the neighborhood set of node v, is the normalization coefficient, is the weight matrix of the lth layer, is the feature vector of node i in layer l, is the activation function. In the management of river basin systems, this hierarchical representation can describe the structural characteristics of the water conservancy system from different scales, including both the local coordinated dispatching relationship of reservoir groups and the global water resources allocation pattern of the basin. After the above processing, the obtained water conservancy system diagram can comprehensively characterize the static connection relationship and dynamic interaction characteristics between water conservancy facilities. Taking the flood control dispatch of a certain river basin as an example, when the upstream reservoir group exceeds the flood limit water level, the diagram representation can quickly identify the downstream sensitive areas, analyze the impact range of different flood discharge plans, and assist in making scientific joint dispatch decisions.

[0067] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0068] (1) The water conservancy data feature vector and the water conservancy system diagram are fused and encoded, and a water conservancy feature fusion matrix is ​​generated through multi-layer cross-attention operations;

[0069] (2) Decouple the water conservancy feature fusion matrix into task-specific features, and obtain the shared feature set and private feature set through symmetric loss constraints and orthogonal decomposition operations;

[0070] (3) Dynamically group the shared feature sets according to the correlation between tasks, and obtain the task combination sequence through mutual information calculation and hierarchical clustering analysis;

[0071] (4) Perform task-specific analysis on the private feature set and obtain the task weight vector by calculating feature importance and ranking contribution;

[0072] (5) Construct a multidimensional prediction target space based on the task combination sequence and task weight vector, and obtain the prediction feature set through adaptive target mapping and constraint optimization;

[0073] (6) The prediction feature set is input into the shared-private feature extractor, and the water conservancy prediction result is obtained through feature recombination and multi-task fusion operations.

[0074] Specifically, the water conservancy data feature vector and the water conservancy system diagram representation are fused. The fusion process is achieved through a multi-layer cross-attention mechanism, which calculates the mutual attention between features at each layer and performs weighted combination. Taking the joint dispatch of a group of reservoirs as an example, the feature fusion matrix contains the characteristic information of the reservoir itself (such as water level, inflow) and the correlation information between reservoirs (such as upstream and downstream relationships, dispatch constraints). The fusion calculation uses the following formula:

[0075] ;

[0076] in, is the feature fusion matrix (unit: dimensionless), is the data feature matrix of the lth layer, is the graph representation matrix of the lth layer, is the weight coefficient of the lth layer, Represents the tensor product operation, and L is the total number of layers. This formula fuses point features and structural features at multiple levels to form a unified feature representation. In the process of feature decoupling, the fused features are separated into shared features and private features through symmetric loss constraints and orthogonal decomposition. For multiple prediction tasks in water conservancy systems, such as flood control scheduling, power generation scheduling, water supply scheduling, etc., each task has specific goals and constraints. The mathematical expression of feature decoupling is:

[0077] ;

[0078] in, is the fusion feature of the i-th task, is the i-th shared feature, is the i-th private feature, are model parameters, is the regularization coefficient, is the number of tasks. The first term ensures the integrity of feature decomposition, and the second term ensures the independence of private features through orthogonal constraints. In practical applications, for example, power generation scheduling tasks mainly focus on water head and unit characteristics, while flood control scheduling focuses more on flood control reservoir capacity and downstream river conditions. Feature decoupling can highlight the core features of different tasks.

[0079] The task combination process is based on mutual information calculation and hierarchical clustering, combining highly related tasks for joint optimization. The combined score is calculated as follows:

[0080] ;

[0081] in, For the i-th task and the jth task The combined score of is the mutual information value of the two tasks, and the numerator represents the i-th shared feature and the jth shared feature The denominator represents the size of the intersection of the two tasks, and the denominator represents the size of the union. This score reflects the degree of correlation and feature sharing between tasks. In the joint dispatching of reservoir groups, flood control and water supply tasks often need to be considered in a coordinated manner during the flood season, while power generation tasks are relatively independent. A reasonable optimization structure can be established through task combination. After feature reorganization and multi-task fusion, multi-dimensional water conservancy prediction results including water level prediction, flow prediction, and dispatching suggestions are generated.

[0082] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0083] (1) The water conservancy forecast results are decomposed hierarchically according to the scheduling objectives, and the scheduling subtask set is obtained through hierarchical constraint analysis;

[0084] (2) Construct the state space of the scheduling subtask set and generate the scheduling decision space through state encoding and action mapping;

[0085] (3) Allocate rewards and penalties in the scheduling decision space according to safety constraints, and obtain the scheduling reward function through multi-dimensional evaluation;

[0086] (4) Perform rolling time domain optimization on the scheduling reward function and obtain the scheduling strategy sequence through predictive control calculation;

[0087] (5) Construct a collaborative scheduling network based on the scheduling strategy sequence, and obtain a scheduling collaboration plan through agent interaction analysis;

[0088] (6) The scheduling coordination plan is input into the hierarchical reinforcement learning framework, and the water conservancy scheduling plan is obtained through strategy optimization.

[0089] Specifically, the water conservancy forecast results are decomposed into layers. In the joint operation of reservoir groups, the overall operation objectives include flood control, power generation, water supply and other aspects. Figure 2 As shown, it is a schematic diagram of the hierarchical decomposition process in the embodiment of the present application. Through hierarchical constraint analysis, the complex scheduling problem is decomposed into subtasks with clear constraints. The flood control subtask needs to consider constraints such as the flood control limit water level and the safe flow of the downstream river; the power generation subtask focuses on parameters such as the head height and the unit output characteristics; the water supply subtask needs to meet the water needs of life, industry and agriculture. The subtask decomposition adopts the following formula:

[0090] ;

[0091] in, is the overall scheduling objective function, is the weight coefficient of the i-th subtask, is the subtask objective function, is the state variable (such as water level, flow), To control variables (such as leakage flow), As constraints, is the number of subtasks. Taking a reservoir group in a river basin as an example, the weight of flood control subtasks is increased during the flood season, while the emphasis is placed on water supply and power generation tasks during the dry season.

[0092] When constructing the state space for each scheduling subtask, it is necessary to discretize and encode continuous variables such as water level and flow, and establish a mapping relationship with specific scheduling actions. The construction formula of the scheduling decision space is:

[0093] ;

[0094] in, For decision space, represents the state vector at time t, S is the state space, is the action vector at time t, For the status The following set of possible actions, In actual dispatching, for example, the discharge of a reservoir needs to consider the gate dispatching capacity and safety constraints, and the continuous discharge flow is discretized into a finite number of dispatching gears.

[0095] Based on the constructed decision space, a scheduling reward function is designed to evaluate the effects of different scheduling strategies. The reward function comprehensively considers multiple dimensions such as flood control safety, power generation efficiency, and water supply guarantee rate:

[0096] ;

[0097] in, is the comprehensive reward function, is the reward value of the kth evaluation dimension, is the corresponding weight, is the penalty for violating the constraint, is the penalty coefficient, and m is the number of evaluation dimensions. In reservoir operation, when the water level exceeds the flood control limit water level or the discharge exceeds the river safety flow, the evaluation score of the operation plan is reduced through the penalty item.

[0098] Through rolling time domain optimization and intelligent collaborative learning, a water conservancy dispatching plan that meets multi-objective constraints is generated. In practical applications, when a river basin encounters a severe flood, the upstream reservoir group adopts measures such as pre-discharge to reduce flood control pressure through coordinated dispatching, while taking into account power generation benefits and ecological water demand, reflecting the comprehensive balance of multi-objective dispatching.

[0099] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0100] (1) The water conservancy dispatching scheme is divided into levels according to the evaluation dimensions, and the evaluation factor set is obtained by indicator decomposition;

[0101] (2) Perform fuzzy rule mapping on the evaluation factor set and obtain the fuzzy feature matrix through membership calculation;

[0102] (3) Carry out association analysis on the fuzzy feature matrix according to knowledge rules and obtain the evaluation rule base through expert experience transformation;

[0103] (4) Learn the indicator weights of the evaluation rule base and obtain the dynamic weight vector through attention calculation;

[0104] (5) Construct a multi-layer evaluation sequence based on the dynamic weight vector and obtain a comprehensive score set through indicator fusion;

[0105] (6) The comprehensive score set is input into the fuzzy neural network, and the water conservancy evaluation results are obtained through multi-dimensional fusion.

[0106] Specifically, the evaluation indicators of water conservancy scheduling schemes are divided into technical dimension, economic dimension, social dimension and environmental dimension. Figure 3As shown, it is a schematic diagram of the evaluation factor set in the embodiment of the present application. The technical dimension includes indicators such as flood control scheduling effect, water supply guarantee rate, power generation benefit, etc.; the economic dimension involves indicators such as project operation cost, power generation income, irrigation benefit, etc.; the social dimension focuses on indicators such as water supply population coverage, agricultural irrigation area, and shipping guarantee; the environmental dimension includes indicators such as ecological base flow satisfaction, water quality compliance rate, and river channel scouring and siltation changes. For the specific indicators under each evaluation dimension, quantification is performed through fuzzy rule mapping. Taking the flood control scheduling effect as an example, it is necessary to comprehensively consider multiple aspects such as reservoir water level control, flood discharge process control, and downstream safety. By setting fuzzy language variables and membership functions, qualitative and quantitative indicators are uniformly converted into fuzzy features. For example, for the reservoir water level control indicator, language variables such as "safe", "general", and "dangerous" are set according to the relationship between the actual water level and the flood control limit water level, and corresponding membership calculation rules are established.

[0107] The fuzzy feature matrix is ​​transformed into an evaluation rule base through knowledge rule transformation. The rule base contains a large number of IF-THEN rules based on expert experience, which describe the evaluation results under different combinations of indicator values. Taking water supply scheduling as an example, when the water supply satisfaction rate is high, the water quality compliance rate is high, and the energy consumption index is low, the water supply scheduling effect is evaluated as "excellent"; when only the basic water supply requirements are met but the energy consumption is high, the evaluation result is "medium". These rules reflect the comprehensive judgment experience of experts on the effect of water conservancy scheduling. Indicator weight learning adopts the attention mechanism to dynamically adjust the weight according to the importance of indicators in different scenarios. In the flood season, the weights of flood control-related indicators will automatically increase; in the dry season, the weights of water supply and ecological related indicators will increase accordingly. By calculating the correlation and importance between indicators, a dynamic weight vector is generated to make the evaluation results better adapt to the actual situation.

[0108] The construction of multi-layer evaluation sequence adopts the hierarchical analysis method. The indicators are integrated within each dimension to obtain the dimension score; then the scores of different dimensions are weighted and combined to obtain the comprehensive score. In the combination process, the mutual influence and constraint relationship between indicators are considered to avoid repeated calculation or ignoring key factors. Finally, the fusion of multi-dimensional evaluation results is realized through fuzzy neural network. The network structure includes fuzzification layer, rule layer, normalization layer and defuzzification layer, which can handle complex nonlinear relationships and has self-learning ability. Through the training of a large number of historical evaluation cases, the network gradually grasps the evaluation rules and improves the accuracy and reliability of the evaluation results.

[0109] Taking the evaluation of a joint dispatching scheme of a certain reservoir group as an example, real-time monitoring data such as water level, flow, water quality, as well as statistical data such as power generation, water supply, and irrigation area are collected. In the technical dimension, the water level control during flood control is analyzed, and the flow compliance rate of each section is calculated; in the economic dimension, the power generation income and operating costs are statistically analyzed; in the social dimension, the water supply and irrigation guarantee rate is evaluated; in the environmental dimension, the ecological flow discharge and water quality change trend are monitored. According to the fuzzy rules set by expert experience, combined with dynamic weight calculation, the comprehensive evaluation results of the dispatching scheme are obtained. The entire evaluation process reflects the organic unity of multi-dimensional indicators, which not only ensures the scientificity and comprehensiveness of the evaluation, but also reflects the dynamic adaptability under different scenarios. In the data processing process, the original monitoring data needs to be preprocessed, including outlier processing, missing value supplementation, data standardization, etc. Then the data is classified and sorted according to the evaluation dimension, and the mapping relationship between the data and the evaluation indicators is established. When mapping fuzzy rules, it is necessary to design a suitable membership function in combination with the characteristics of the indicators to achieve qualitative to quantitative transformation. The extraction and sorting of knowledge rules is an iterative process, which requires continuous accumulation and summary of expert experience. Weight learning involves the training of a large amount of historical data, and attention should be paid to the representativeness and completeness of the samples. In the process of multi-layer evaluation, special attention should be paid to the logical relationship between indicators to ensure the rationality of the evaluation results.

[0110] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0111] (1) Analyze the water conservancy evaluation results into data streams and obtain real-time data streams through continuous monitoring;

[0112] (2) Perform incremental feature analysis on real-time data streams and obtain updated feature sets through data change detection;

[0113] (3) Segment the updated feature set according to the time series rules and obtain the data segment sequence through dynamic window division;

[0114] (4) Perform concept drift detection on the data segment sequence and obtain the update trigger signal by calculating the distribution difference;

[0115] (5) Construct an incremental learning sample set based on the update trigger signal, and obtain a high-quality sample library through sample value evaluation;

[0116] (6) Input the high-quality sample library into the data stream processing engine and obtain the optimized water conservancy service model through progressive training.

[0117] Specifically, water conservancy data stream processing requires data stream analysis of water conservancy evaluation results. Data stream analysis includes real-time monitoring data of multiple dimensions, such as water level data, flow data, water quality data, etc. Continuous monitoring obtains these monitoring data in real time through the data acquisition terminal and standardizes them according to a unified data format. After cleaning, denoising and format conversion, the original data forms a standardized real-time data stream. Each data record contains fields such as timestamp, monitoring point identifier, monitoring indicator and monitoring value. Incremental feature analysis performs feature extraction and change detection on real-time data streams. The newly arrived data is compared with historical data to calculate the change trend and fluctuation range of each indicator. Feature analysis includes basic statistical features (such as mean, variance, extreme value, etc.), time series features (such as periodicity, trend, etc.) and correlation features (such as the relationship between indicators). Data change detection focuses on the mutation points of indicator values, the abnormality of change rate and the change of correlation between multiple indicators. The updated feature set is obtained through the analysis of these features.

[0118] The data segmentation process uses a dynamic window mechanism to divide the update feature set into multiple data segments according to the timing rules. The window size is dynamically adjusted according to the characteristics of the data. Smaller windows are used for faster-changing data, and larger windows are used for slower-changing data. Each data segment contains complete timing information and feature descriptions to facilitate subsequent analysis and processing. A certain overlap is maintained between data segments to ensure that important change features are not missed. The step size of the window sliding is also dynamically adjusted. When the data changes drastically, the step size is shortened, and when the change is stable, the step size is increased. Concept drift detection identifies model performance degradation by analyzing distribution changes in the sequence of data segments. Distribution difference calculation includes multiple aspects such as the shift of feature distribution, changes in data correlation, and the increase of model prediction error. When a significant distribution difference is detected, an update trigger signal is generated to start the model update process. Concept drift detection needs to consider both sudden changes and gradual changes, and adopt different detection strategies for different types of changes.

[0119] The construction of the incremental learning sample set is based on the update trigger signal, and representative samples are screened from the new data. The sample value assessment takes into account the three aspects of information volume, representativeness and diversity. Samples with high information volume usually contain new data patterns or boundary conditions; samples with strong representativeness can reflect the main distribution characteristics of the data; and diversity ensures that the samples cover different scenarios and conditions. Through the evaluation of these standards, high-quality samples are selected to be added to the sample library. The data stream processing engine adopts a progressive training strategy and continuously optimizes the model using a high-quality sample library. Progressive training not only maintains the model's memory of historical knowledge, but also can learn new data patterns. During the training process, the learning rate is dynamically adjusted, and different weights are given to new samples and historical samples to ensure the stability of model updates.

[0120] Taking the joint dispatch of a group of reservoirs as an example, the real-time monitoring system collects water level, flow and other data every 5 minutes. Data stream parsing converts monitoring data from different sources into a standard format, including timestamp, monitoring point ID, indicator type and measurement value. Incremental feature analysis found that the water level rise rate of a certain reservoir was abnormal. By calculating the statistical characteristics of historical data of the same period, it was determined that this was a change feature that needed attention. The dynamic window was automatically adjusted to 30 minutes according to the rate of change of the water level, generating a series of data fragments reflecting the rapid change of the water level. Concept drift detection found that the current dispatch model had deviations in handling such rapid rises, triggering a model update signal. Similar rapid rise cases were screened from recent data to construct an incremental learning sample set. After value assessment, these samples were added to the high-quality sample library for progressive training of the model to improve the model's ability to respond to rapid water level changes.

[0121] The above describes the method for constructing a water conservancy big data service analysis and evaluation model in the embodiment of the present application. The following describes the system for constructing a water conservancy big data service analysis and evaluation model in the embodiment of the present application. Figure 4 In the embodiment of the present application, an embodiment of the water conservancy big data service analysis and evaluation model construction system includes:

[0122] The extraction module is used to extract multimodal features of water conservancy monitoring data through a two-stream attention mechanism to obtain a water conservancy data feature vector;

[0123] A generation module, used for constructing a dynamic heterogeneous graph neural network according to the water conservancy data feature vector, and generating a water conservancy system graph representation through an adaptive graph convolution operation;

[0124] A prediction module, used to use the water conservancy data feature vector and the water conservancy system diagram representation to perform multi-task prediction on the water conservancy index through a shared-private feature extractor to obtain a water conservancy prediction result;

[0125] A calculation module, used to optimize the water conservancy dispatching through a hierarchical reinforcement learning framework based on the water conservancy prediction results to generate a water conservancy dispatching plan;

[0126] An analysis module is used to conduct a comprehensive analysis of the water conservancy system through a fuzzy neural network according to the water conservancy dispatching plan to obtain a water conservancy evaluation result;

[0127] The updating module is used to perform incremental updates through a data stream processing engine according to the water conservancy evaluation results to obtain an optimized water conservancy service model.

[0128] Through the collaboration of the above components, the multimodal feature extraction of water conservancy monitoring data is carried out through the dual-stream attention mechanism, which effectively solves the problem of unified processing of structured data and unstructured data, and realizes the deep integration of temporal features and spatial features. Based on the extracted water conservancy data feature vector, the dynamic heterogeneous graph neural network is used to construct the overall representation of the water conservancy system. The dynamic correlation between water conservancy facilities is effectively captured through adaptive graph convolution operation, which enhances the ability to express the structural characteristics of complex water conservancy systems. The introduction of shared-private feature extractors for multi-task prediction not only ensures the sharing and transfer of knowledge between different prediction tasks, but also maintains the specific characteristics of each task, thereby improving the prediction accuracy. The hierarchical reinforcement learning framework is used for water conservancy scheduling optimization, which decomposes the complex scheduling problem into multiple subtasks, and realizes the global optimal scheduling through the collaborative learning of intelligent agents. The application of fuzzy neural networks enables the evaluation process to process qualitative and quantitative indicators at the same time, which improves the scientificity and interpretability of the evaluation results. The data stream processing engine realizes the online learning and continuous optimization of the model, and enables the model to adapt to the dynamically changing water conservancy environment through the incremental update mechanism.

[0129] Reference Figure 5 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0130] Those skilled in the art will understand that Figure 5 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0131] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0132] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0133] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0134] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0135] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for constructing a water conservancy big data service analysis and evaluation model, characterized in that: The water conservancy big data service analysis and evaluation model construction method includes: The multimodal features of water conservancy monitoring data are extracted through the two-stream attention mechanism to obtain the water conservancy data feature vector; A dynamic heterogeneous graph neural network is constructed according to the water conservancy data feature vector, and a water conservancy system graph representation is generated through an adaptive graph convolution operation, including: constructing a water conservancy node set according to the spatial position relationship of the water conservancy data feature vector, and obtaining a water conservancy edge set through correlation calculation; encoding the physical relationship between the water conservancy node set and the water conservancy edge set to generate an initial graph topology structure; inputting the initial graph topology structure into a time-varying graph analyzer, and obtaining a graph evolution sequence through dynamic connection strength calculation; generating a multi-level graph representation based on the graph evolution sequence, and obtaining a water conservancy sub-graph set through hierarchical aggregation operation; performing structural comparative analysis on the water conservancy sub-graph set, and obtaining a graph matching matrix through topological similarity calculation; inputting the graph matching matrix into an adaptive graph convolution operation to generate a water conservancy system graph representation; Utilizing the water conservancy data feature vector and the water conservancy system diagram representation, a shared-private feature extractor is used to perform multi-task prediction on the water conservancy index to obtain a water conservancy prediction result, including: information fusion encoding of the water conservancy data feature vector and the water conservancy system diagram representation, generating a water conservancy feature fusion matrix through multi-layer cross attention operations; performing task feature decoupling on the water conservancy feature fusion matrix, obtaining a shared feature set and a private feature set through symmetric loss constraints and orthogonal decomposition operations; dynamically grouping the shared feature set according to the correlation between tasks, obtaining a task combination sequence through mutual information calculation and hierarchical clustering analysis; performing task specificity analysis on the private feature set, obtaining a task weight vector through feature importance calculation and contribution ranking; constructing a multidimensional prediction target space based on the task combination sequence and task weight vector, obtaining a prediction feature set through adaptive target mapping and constraint optimization; inputting the prediction feature set into a shared-private feature extractor, and obtaining a water conservancy prediction result through feature recombination and multi-task fusion operations; Based on the water conservancy prediction results, the water conservancy dispatching is optimized and calculated through a hierarchical reinforcement learning framework to generate a water conservancy dispatching plan; According to the water conservancy dispatching plan, a comprehensive analysis of the water conservancy system is performed through a fuzzy neural network to obtain a water conservancy evaluation result; According to the water conservancy evaluation results, incremental updates are performed through a data stream processing engine to obtain an optimized water conservancy service model.

2. The method for constructing a water conservancy big data service analysis and evaluation model according to claim 1, characterized in that: The multimodal feature extraction of water conservancy monitoring data is performed through the dual-stream attention mechanism to obtain the water conservancy data feature vector, including: The water conservancy monitoring data is processed by splitting the data through the time series convolutional network to generate structured data streams and unstructured data streams; Input the structured data stream into the temporal attention module, and obtain a temporal feature matrix through multi-scale feature decomposition; Inputting the unstructured data stream into a spatial attention module, and obtaining a spatial feature matrix by adaptively partitioning the region; Performing a cross attention operation on the temporal feature matrix and the spatial feature matrix to obtain a fused feature tensor; Constructing contrastive learning sample pairs based on the fused feature tensor, and obtaining a positive sample set through data enhancement; The positive sample set is input into the contrast loss function for feature mapping to obtain a water conservancy data feature vector.

3. The method for constructing a water conservancy big data service analysis and evaluation model according to claim 1, characterized in that: Based on the water conservancy prediction results, the water conservancy dispatching is optimized and calculated through a hierarchical reinforcement learning framework to generate a water conservancy dispatching plan, including: Decomposing the water conservancy forecast results in layers according to the scheduling objectives, and obtaining a scheduling subtask set through hierarchical constraint analysis; Constructing a state space for the scheduling subtask set, and generating a scheduling decision space through state encoding and action mapping; The scheduling decision space is assigned rewards and penalties according to safety constraints, and a scheduling reward function is obtained through multi-dimensional evaluation; Performing rolling time domain optimization on the scheduling reward function, and obtaining a scheduling strategy sequence through predictive control calculation; Building a collaborative scheduling network based on the scheduling strategy sequence, and obtaining a scheduling collaborative solution through agent interaction analysis; The scheduling coordination plan is input into the hierarchical reinforcement learning framework, and the water conservancy scheduling plan is obtained through strategy optimization.

4. The method for constructing a water conservancy big data service analysis and evaluation model according to claim 1, characterized in that: According to the water conservancy dispatching plan, the water conservancy system is comprehensively analyzed through a fuzzy neural network to obtain a water conservancy evaluation result, including: Divide the water conservancy dispatching plan into levels according to the evaluation dimensions, and obtain the evaluation factor set by index decomposition; Performing fuzzy rule mapping on the evaluation factor set, and obtaining a fuzzy feature matrix by calculating the membership degree; Performing association analysis on the fuzzy feature matrix according to knowledge rules, and obtaining an evaluation rule base through expert experience transformation; Performing indicator weight learning on the evaluation rule base, and obtaining a dynamic weight vector through attention calculation; Constructing a multi-layer evaluation sequence based on the dynamic weight vector, and obtaining a comprehensive score set by integrating indicators; The comprehensive score set is input into the fuzzy neural network, and the water conservancy evaluation result is obtained through multi-dimensional fusion.

5. The method for constructing a water conservancy big data service analysis and evaluation model according to claim 1, characterized in that: The method of performing incremental updates based on the water conservancy evaluation results through a data stream processing engine to obtain an optimized water conservancy service model includes: Performing data stream analysis on the water conservancy evaluation results and obtaining real-time data streams through continuous monitoring; Performing incremental feature analysis on the real-time data stream, and obtaining an updated feature set through data change detection; Segment the updated feature set according to the time series rule, and obtain a data segment sequence through dynamic window division; Performing concept drift detection on the data segment sequence, and obtaining an update trigger signal by calculating distribution differences; Building an incremental learning sample set based on the update trigger signal, and obtaining a high-quality sample library through sample value evaluation; The high-quality sample library is input into the data stream processing engine, and an optimized water conservancy service model is obtained through progressive training.

6. A water conservancy big data service analysis and evaluation model construction system, used to implement the water conservancy big data service analysis and evaluation model construction method as described in any one of claims 1 to 5, characterized in that: The water conservancy big data service analysis and evaluation model construction system includes: The extraction module is used to extract multimodal features of water conservancy monitoring data through a two-stream attention mechanism to obtain a water conservancy data feature vector; A generation module, used for constructing a dynamic heterogeneous graph neural network according to the water conservancy data feature vector, and generating a water conservancy system graph representation through an adaptive graph convolution operation; A prediction module, used to use the water conservancy data feature vector and the water conservancy system diagram representation to perform multi-task prediction on the water conservancy index through a shared-private feature extractor to obtain a water conservancy prediction result; A calculation module, used to optimize the water conservancy dispatching through a hierarchical reinforcement learning framework based on the water conservancy prediction results to generate a water conservancy dispatching plan; An analysis module is used to conduct a comprehensive analysis of the water conservancy system through a fuzzy neural network according to the water conservancy dispatching plan to obtain a water conservancy evaluation result; The updating module is used to perform incremental updates through a data stream processing engine according to the water conservancy evaluation results to obtain an optimized water conservancy service model.

7. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, it implements the water conservancy big data service analysis and evaluation model construction method described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor executes the method for constructing a water conservancy big data service analysis and evaluation model as described in any one of claims 1 to 5.

Images