Intelligent flood scheduling scheme recommendation method based on scheduling business knowledge graph

Abstract

The invention discloses a flood intelligent scheduling scheme recommendation method based on a scheduling business knowledge graph, and the method comprises the steps: 1, constructing the scheduling business knowledge graph, including entity recognition, knowledge extraction and knowledge fusion; step 2, hydrological time sequence anomaly monitoring: detecting and correcting an abnormal value in forecast data; step 3, constructing a flood scheduling auxiliary decision based on knowledge graph pattern reasoning, finding out more than one rule of which the premise is matched with facts or closest to facts to form an available rule set by adopting a relaxation matching method, and obtaining a reservoir scheduling strategy by adopting a weighted voting mode; the problems that in the prior art, flood dispatching response is not timely, dispatching strategies are insufficient, and accuracy and real-time performance of flood dispatching are poor are solved.

Description

Technical Field

[0001] This invention belongs to the field of flood control technology; in particular, it relates to a method for recommending intelligent flood control schemes based on a knowledge graph of control operations. Background Technology

[0002] In actual hydropower station operation, the role of dispatching regulations is often overlooked. This is partly due to the inefficiency of reading these regulations and the relatively high level of familiarity among workers. Firstly, dispatching regulations typically contain a large amount of complex information, requiring considerable time and effort to understand and master. This adds an extra burden to busy dispatchers, leading them to rely more on existing experience and skills. Secondly, in practice, particularly regarding flood control, the lack of a systematic knowledge base of dispatching experts and the ineffective accumulation and transmission of relevant dispatching experience result in delayed responses, inadequate dispatching strategies, and poor accuracy and real-time performance in flood control operations. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a method for recommending intelligent flood control scheduling schemes based on a scheduling business knowledge graph, so as to solve the problems of untimely response, inadequate scheduling strategies, and poor accuracy and real-time performance of existing flood control scheduling technologies.

[0004] The technical solution of this invention is:

[0005] A method for recommending intelligent flood control scheduling schemes based on a scheduling business knowledge graph, the method comprising:

[0006] Step 1: Construct a knowledge graph based on scheduling services, including entity recognition, knowledge extraction, and knowledge fusion;

[0007] Step 2, Hydrological Time Series Anomaly Monitoring: Detecting and correcting outliers in the forecast data;

[0008] Step 3: Construct a flood control scheduling auxiliary decision-making system based on knowledge graph graph pattern reasoning to find a correct scheduling implementation plan.

[0009] The entity recognition method includes: First, a label set is formulated based on the corpus samples. Each element in the sample sequence is labeled according to the water conservancy features. Based on this, the preprocessed unstructured water conservancy text is labeled to determine the boundaries of words in the text and to determine the classification features for entity recognition. Then, a BiLSTM model is trained. Sentence features are extracted by saving the context information of the whole sentence, and the forward and backward hidden state results in the text are combined. Finally, the output of the BiLSTM model is passed to CRF as input to form a BiLSTM-CRF structure, and the sequence is labeled using context information. In the model training, a method based on word vector combination is used. An automatically acquired dictionary is used to match sentences, and the input character sequence and all latent words are compiled.

[0010] The knowledge extraction methods include: using a combination of pattern matching and co-occurrence network analysis to extract relationships between water conservancy entities. The pattern matching method uses linguistic knowledge to construct entity relationship patterns based on word, part-of-speech, or semantic features of water conservancy entities, and then matches the water conservancy entity text with the patterns to extract relationships. The co-occurrence network analysis method determines the relationships between entities by constructing a co-occurrence matrix. First, it uses statistical methods to extract the frequency of each water conservancy entity in the text, and then analyzes the co-occurrence ratio of different entities in the text. When the co-occurrence ratio of two entities is greater than a certain threshold, it is considered that there is a relationship between the two entities. The relationship between entities is defined based on the type of entity.

[0011] Methods of knowledge fusion include:

[0012] Entity fusion: First, the similarity between two strings is calculated using the edit distance metric. Then, the similarity of related attributes is analyzed based on the attribute information of the two entities. Finally, the graph structure similarity is calculated based on the graph relationship between the two entities. Through comprehensive analysis of the three similarities, if the similarity exceeds a set threshold, the two entities are fused.

[0013] Cross-domain fusion of maps: The fusion of water network maps and water conservancy discipline maps includes the fusion of water conservancy principles or laws with water network entities, and the fusion of knowledge between conceptual terms and water network entities; first, entity alignment technology is used to match the local features of entity characters, and then global features are used to match related entities in the two maps.

[0014] The method for detecting outliers is to use an improved K-means clustering algorithm for time series data.

[0015] Constructing flood control and dispatching auxiliary decision-making based on knowledge graph schema reasoning includes:

[0016] Step 3.1: Knowledge Graph Pattern Reasoning; By extracting patterns from the knowledge graph and performing reasoning, the relationships and patterns hidden in the data are discovered.

[0017] Step 3.2, Flood Control Scheme Assisted Decision Making: Use inference algorithms to determine the flood control strategy.

[0018] The knowledge graph pattern reasoning method includes:

[0019] Pattern extraction: Extracting patterns related to queries or knowledge from knowledge graphs. Using graph mining algorithms to find subgraphs that appear frequently in the graph, these subgraphs are considered to represent patterns of a specific knowledge or query.

[0020] Pattern matching: Finding entities, relationships, and attributes in a knowledge graph that match a query or knowledge, using graph matching algorithms, rule-based algorithms, or graph traversal-based algorithms.

[0021] Methods for determining flood control strategies using inference algorithms include: using a relaxed matching method to find one or more rules that match or are closest to the facts to form a set of available rules, and using a weighted voting method to obtain the reservoir's control strategy.

[0022] The beneficial effects of this invention are:

[0023] This invention performs knowledge extraction, knowledge fusion, and knowledge reasoning operations on text files such as reservoir management and scheduling procedures and scheduling logs. It adopts a top-down approach to build a model layer, data layer, technology layer, and application layer, thereby realizing the intelligent processing, representation, and utilization of hydrological information.

[0024] This invention establishes an anomaly identification system, employs a flood prediction algorithm to identify outliers and reduce noise in monitoring data, thereby separating actual anomalies and providing hydrological data monitoring and alarm functions.

[0025] This invention uses a constructed flood control knowledge graph to optimize flood control schemes while ensuring the universality and stability of the model, thereby improving the accuracy and real-time performance of flood control.

[0026] It solves the problems of untimely response, inadequate scheduling strategies, and poor accuracy and real-time performance of existing flood control technologies. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the knowledge graph design process of the present invention;

[0028] Figure 2 This is a schematic diagram of the knowledge graph structure of the present invention;

[0029] Figure 3This is a schematic diagram of the outlier detection process for flood time series based on the K-means clustering algorithm. Detailed Implementation

[0030] A method for recommending intelligent flood control scheduling schemes based on a scheduling business knowledge graph includes:

[0031] Step 1: Knowledge Graph Design

[0032] Data layer: Collects and classifies source data, including structured data exported from relational databases in the water conservancy industry, semi-structured web page data, and unstructured text data, serving as the data foundation for extracting water conservancy entities and relationships;

[0033] Technical Layer: Relevant entities are extracted from text or image data using natural language processing, machine learning, and other technologies. These entities are then categorized into natural objects, engineering objects, and social objects. Relationships between different entities are extracted using a predefined water network entity relationship model. An encyclopedic graph is constructed based on the classification of water conservancy disciplines and the relationships between terminology concepts. Furthermore, multi-source entity and cross-domain graph fusion is performed, including ontology alignment, semantic association and attribute merging based on multi-source entities, and the integration of water conservancy knowledge models. Finally, the graph is represented in a structured form and stored using a graph database. This invention employs Neo4j, a widely used graph database. Data storage includes two basic types: nodes and relationships. Nodes are connected through defined relationships to form a relational network structure.

[0034] Application layer: Based on the constructed knowledge graph, it can first provide applications such as visualization and related queries, and also facilitate the clustering of different water conservancy objects; in addition, with the expansion of water conservancy entities and the enrichment of entity relationships, it can also provide knowledge applications such as graph-based knowledge question answering and tracing the causes of water conservancy events.

[0035] Model layer: Based on the scheduling measures and objectives of flood control, a conceptual model is designed and formed from aspects such as hierarchy, entity type and attributes and relationships between entities, and the model layer is constructed.

[0036] The construction of a flood control knowledge graph mainly includes entity recognition, knowledge extraction, and knowledge fusion.

[0037] Entity Recognition: In knowledge graphs, independently existing things (objects or concepts) are called entities, and the basic unit of the knowledge graph is formed by "entity-relationship-entity" triples. For structured water conservancy entities, relational data can be directly converted into triples based on data tables. When recognizing water conservancy entities in unstructured text, entity recognition needs to be transformed into a sequence labeling problem. Therefore, a method combining Bi-directional Long Short-Term Memory Neural Network (BiLSTM) and Conditional Random Fields (CRF) is chosen to recognize water conservancy entities, and a method based on word vectors is selected for word segmentation. The BiLSTM model has powerful nonlinear fitting and sequence modeling capabilities, and can capture long-term contextual information. CRF is a statistical learning model commonly used for labeling problems, especially with small datasets, and has good recognition performance. Combining BiLSTM and CRF for named entity recognition can fully leverage their respective advantages. The BiLSTM model is used to mine feature information in the text, and the results are then input into the CRF model for sequence labeling.

[0038] The corpus for identifying water conservancy entities in this invention mainly consists of unstructured descriptions of water network objects. First, a labeling set was established based on the corpus samples. Each element in the sample sequence was labeled according to water conservancy features, and based on this, the preprocessed unstructured water conservancy text was labeled to determine word boundaries and establish classification features for entity recognition. Then, a BiLSTM model was trained, extracting sentence features by preserving the context information of the entire sentence and combining the forward and backward hidden state results from the text. Finally, the output of the BiLSTM model was passed to a CRF as input, forming a BiLSTM-CRF structure, and sequence labeling was performed using context information. The model training used a method based on combining character and word vectors, employing an automatically acquired dictionary to match sentences and compiling the input character sequence and all latent words.

[0039] Knowledge Extraction: The goal is to extract relationships between entities from unstructured text data. This invention employs a combination of pattern matching and co-occurrence network analysis to extract relationships between water conservancy entities. Pattern matching utilizes linguistic knowledge to construct entity relationship patterns based on word, part-of-speech, or semantic features of water conservancy entities, and then matches the water conservancy entity text with these patterns for relationship extraction. During the extraction process, the expression of relationships between entities is first constructed based on water conservancy corpus samples. For example, to express the "inclusion" relationship between a river basin and a power station, a relationship pattern of [river basin name] including [power station name] can be constructed. The co-occurrence network analysis method determines relationships between entities by constructing a co-occurrence matrix. Its basic assumption is that closely related entities will appear simultaneously in multiple segments of the text. First, the frequency of each water conservancy entity in the text is extracted using statistical methods. Then, the co-occurrence ratio of different entities in the text is analyzed. When the co-occurrence ratio of two entities exceeds a certain threshold, a relationship is considered to exist between the two entities. Relationships between entities can be defined based on entity types, such as a relationship between a river and a region (e.g., a relationship of flow through the river), or a relationship of jurisdiction between an institution and a river.

[0040] Knowledge integration: This mainly involves unifying the descriptions of the same entity or concept from different data sources, enabling heterogeneous knowledge graphs to communicate with each other.

[0041] Entity fusion addresses the lack of unified identification for the same entity across different data sources by constructing semantic relationships between entities from different data sources. Attribute fusion resolves the consistency issue of attributes for the same entity across different data sources. Disambiguation of entities with the same name involves fusing entities in water conservancy objects that have the same name but different meanings, different names but the same meaning, or the same name from multiple sources (such as place names and reservoir names), primarily based on matching entity structure and attribute features. In disambiguating water conservancy entities, all data nodes are projected into a global unified representation space. A local representation function is constructed for each candidate set to measure the similarity between two entity nodes. Then, nodes are clustered based on similarity, and the candidate sets are split according to the clustering results to obtain the final data fusion result.

[0042] Taking entity fusion as an example, the entity "Reservoir and Dam Center" is extracted from the procedure document, while the organizational data imported from the meeting minutes contains the entity "Centralized Control Side." These two entity names describe the same entity, but because the data sources are different, they need to be merged. First, the similarity between the two strings is calculated using the edit distance index. Then, the similarity of relevant attributes is analyzed based on the attribute information of the two entities. Finally, the graph structure similarity is calculated based on the graph relationship between the two entities (such as hierarchical organizational relationships). By comprehensively analyzing these three similarity metrics, if the result exceeds a set threshold, the two entities are merged.

[0043] Cross-domain fusion of maps refers to the integration of water network maps and water conservancy discipline maps, including the integration of water conservancy principles or laws with water network entities, and the knowledge integration between conceptual entries and water network entities. Technically, entity alignment technology is first used to match local features of entity characters, and then global features are used to match related entities in the two maps. Regarding scholar fusion, the main focus is on matching and updating newly added scholars with existing experts in the database. The system uses scholar attribute information to determine if there is overlap with experts in the database, and resolves the consistency issue for duplicate scholars. In terms of the integration of keywords and water network entities, this is achieved by establishing connections between keywords and water network entities, such as constructing relationships like "discipline field—researcher—institution—research object," thus linking subject keywords with specific people, institutions, rivers, and other objects within the water network. The cross-domain fusion of the flood control project object "Beipanjiang River Basin" and the encyclopedia entry "early warning" is used as an example for illustration. When performing cross-domain fusion, the first step is to search for co-occurrence relationships between the two based on the data collected by the platform, including co-occurrence in texts such as papers and descriptions in expert research fields. Then, the fusion is performed based on the co-occurrence situation. If a co-occurrence relationship exists and the carrier of co-occurrence (such as a paper) is an entity in the established graph, the two are linked through the intermediate entity of co-occurrence. If the carrier of co-occurrence is not in the existing graph, the method of co-occurrence network analysis is used to determine whether to establish a relationship between the two.

[0044] Step 2: Monitoring of hydrological time series anomalies:

[0045] The accuracy of hydrological data is crucial to the correctness of flood control instructions. To ensure the effectiveness of the control, it is essential to ensure that the input data is free of outliers. This step involves detecting and correcting outliers in the forecast data based on the real-time hydrological situation.

[0046] Step 2.1, Hydrological Information Prediction:

[0047] To better represent time series data, it is necessary to reset and filter the time series data for simulation. This invention combines wavelet transform with gene expression to improve the prediction of time series data and improve the final detection accuracy.

[0048] First, the input time series data is processed by DWT, which uses its efficiency to decompose the input time series data into approximate components and detail components. However, not all the sub-signals obtained after decomposition provide the same information.

[0049] Secondly, the sub-signals of the detail components are input into MGGP to eliminate redundant information that may interfere with model performance. MGGP generates an explicit expression representing the contribution to the objective, and MGGP is then used as a tool for selecting expanded details. Because it linearly combines GP trees and only improves fitness based on the objective function, it often produces complex evolutionary solutions, which may lead to horizontal inflation and overfitting. To address this issue, the Pareto front (PF) method is used to optimize the model's accuracy and complexity. The specific process is as follows: first, the maximum number of genes is limited to reduce model complexity and the risk of overfitting; then, the complexity of all evolutionary solutions is calculated, and the overall complexity and fit of the model are compared; finally, the accuracy and complexity of the model are balanced, and the scaled detail components in the best model are selected.

[0050] Finally, the approximate components and scaled detail components are used as external inputs to the nonlinear autoregressive network. This dynamic network consists of three layers (similar to MLPNN). This network has a feedback function, forming a feedback loop by connecting the previous outputs, giving it multi-step prediction characteristics. The specific calculation process is as follows:

[0051]

[0052] In the formula, y, a, and ds are the output, approximation component, and scaling detail component, respectively; Δt1 is the input delay; Δt2 is the feedback delay; and F is the function sign.

[0053] The delayed y is automatically fed back to the input layer. Therefore, the prediction depends not only on the external parameters (a and ds) but also on the previous output signal. At the same time, the delay between the external input and the feedback input cuts off the network's short-term memory. When the network is in a closed loop, it may reduce computational efficiency and accuracy. However, for a known and fixed training set, the true output is valid. Therefore, the loop can be opened by the feedback delay mechanism. With the update and iteration of the input, better predictions can be obtained, and computational efficiency is also improved.

[0054] Step 2.2, Flood Time Series Outlier Detection:

[0055] In the process of checking for outliers in time series data, this invention employs an improved K-means clustering algorithm specifically designed for time series data.

[0056] By extracting statistical and structural features from the dataset, dimensionality reduction is further performed on the subsequences segmented by the sliding window. To construct three-dimensional vectors from the extracted feature values, this invention extends the traditional K-means clustering method to three-dimensional space for clustering time-series data. Addressing the drawback of requiring manual specification of the k-value in the traditional K-means method, the AIC criterion is introduced to test the data and determine the optimal k-value. The AIC criterion is a standard for measuring the goodness of fit of a statistical model; it is based on the concept of entropy and can balance the complexity of the estimated model with its goodness of fit to the data. Combining the classic K-means clustering algorithm, the dataset after dimensionality reduction of the subsequences segmented by the sliding window is obtained. Subsequently, an improved K-means clustering method was developed for time-series data. The dimensionality-reduced dataset obtained in this paper... each All are three-dimensional vectors, given value Under these conditions, the present invention divides the original data into Class, that is Then, find the minimum value of the subset center, as shown in the following formula:

[0057]

[0058] in Representing categories average

[0059] Because hydrological time-series data possesses both duration and randomness, the clusters of hydrological data tend to be relatively tightly bound. Now, let's assume we are performing correlation anomaly detection on the hydrological time-series data of two stations, given the following time-series data:

[0060] M site hydrological time series dataset

[0061]

[0062]

[0063] Let n be the data clusters of this station.

[0064] N-site hydrological time series dataset:

[0065]

[0066]

[0067] Let n be the data clusters of this station.

[0068] After clustering the data from stations M and N using the improved K-means algorithm, their specific cluster distributions are obtained. For ease of discussion, it is assumed that the clustering result for station M is two clusters: , The clustering results for station N are as follows: , Assuming The distribution location is in Similar locations; and Similar locations. When displaying clustering results, based on... The data is used to reconstruct the corresponding time interval, and then the time interval is traced back to... The position in the cluster, if it appears Within a cluster, it can be determined that it is not abnormal. However, this cannot guarantee that... All the data in the restored time interval happened to appear Within the cluster, there is some reason to suspect. The cluster exhibits anomalies. At this point, the concept of confidence intervals is introduced: According to the "Hydrological Survey Specifications," the reliability requirement for hydrological station data monitoring is 95%. If 95% of the data in this dataset is normal, then it is clear that the remaining 5% of the data has significant anomalies. Therefore, if... After cluster data backtracking If the probability of an object in a cluster is greater than the set confidence level P, then the anomaly is confirmed by judging the distance between the object in that cluster and the central object. The first 95% of the data is taken as normal data, and the rest... In clusters and Data within a cluster is considered anomalous; conversely, if data is not found in a cluster, it is assumed that a certain proportion of normal data exists within similar clusters. This method is used to sequentially verify the data. , The two stations were backtracked to... , Anomaly detection at both stations. Anomaly detection based on spatial relationships is mutual. Therefore, while performing anomaly detection on station N data using data from station M, it is also essential to consider the results of anomaly detection on station M data using data from station N. Comparative analysis is necessary to arrive at a comprehensive and reasonable detection result.

[0069] Step 3: Flood control decision support based on knowledge graph schema reasoning:

[0070] Step 3.1, Knowledge Graph Pattern Reasoning: Knowledge graph-based graph pattern reasoning is a method that extracts patterns from a graph and performs reasoning to discover hidden relationships and patterns in data. In graph pattern reasoning, a pattern is a template composed of a set of entities, relations, and attributes. These patterns can be used to represent specific knowledge or queries.

[0071] The reasoning process includes:

[0072] Pattern extraction phase: Extracting patterns related to queries or knowledge from the knowledge graph. This can be achieved using graph mining algorithms, text mining algorithms, or a combination of methods. For example, frequent subgraph mining algorithms can be used to find subgraphs that appear frequently in the graph; these subgraphs can be considered as patterns representing specific knowledge or queries.

[0073] Frequent subgraph mining algorithms can be implemented in various ways, such as graph compression algorithms, search-based algorithms, and frequent itemset-based algorithms. These algorithms are widely used in knowledge graph-based graph pattern reasoning, helping to discover hidden patterns and relationships in data and gain deeper understanding and insights.

[0074] Frequent subgraph mining is a graph mining algorithm designed to find frequently occurring subgraphs from a given set of graphs. This algorithm can be used in graph pattern reasoning based on knowledge graphs to find patterns related to queries or knowledge. Below are the basic steps of the frequent subgraph mining algorithm:

[0075] Choosing a threshold for subgraph size: First, a threshold needs to be set to represent the minimum subgraph size. Only when the size of a subgraph exceeds this threshold will it be considered as a frequent subgraph.

[0076] Generating all candidate subgraphs: For a given set of graphs, a method similar to the Apriori algorithm can be used to progressively generate all possible subgraphs, and then filter out subgraphs that do not meet the requirements based on a threshold. For example, a method similar to Depth-First Search (DFS) can be used, starting from a starting node and continuously expanding to the next level of nodes until the threshold size is reached or expansion is no longer possible.

[0077] Calculate the support for each candidate subgraph: For all candidate subgraphs, we need to calculate their frequency or support in the graph set. Generally, support can be defined as the proportion of graphs that satisfy the subgraph condition out of the total number of graphs.

[0078] Filter out all frequent subgraphs: Filter out all subgraphs whose support is greater than the support threshold based on the support threshold. These subgraphs are the frequent subgraphs.

[0079] Further optimization: Some algorithms further optimize the selected frequent subgraphs, such as removing subgraphs with containment relationships and merging similar subgraphs, to reduce storage and computation costs.

[0080] Pattern matching phase: This phase involves finding entities, relationships, and attributes in the knowledge graph that match the query or knowledge. This is achieved through graph matching algorithms, rule-based algorithms, or graph traversal-based algorithms. For example, subgraph matching algorithms can be used to find subgraphs that match the query or knowledge, or rule-based algorithms can be used to apply predefined rules to derive conclusions relevant to the query or knowledge.

[0081] Subgraph matching algorithms are used to find matching subgraphs within a larger graph. They are commonly used in graph pattern reasoning based on knowledge graphs. The core idea is to compare the structures of two graphs to determine if a matching relationship exists. These algorithms can be implemented using VF2, Ullmann, RI, or LAD algorithms.

[0082] Step 3.2, Flood Control Scheme Assistance Decision Making:

[0083] Using the knowledge graph pattern reasoning method described above, and then establishing an analysis path based on conditions, measures, and target factors to analyze the scheduling scheme, we can achieve auxiliary decision-making recommendation.

[0084] Flood control decision-making schemes are obtained through rule set knowledge reasoning. If the best scheme is to be selected from several schemes as the final control decision, then further optimization of feasible schemes is required. Typically, the merits of a scheme are described by the characteristic values ​​of several evaluation indicators. Scheme evaluation should be based on the evaluation criteria or indicators corresponding to the control objectives, using a multi-objective decision-making fuzzy optimization method to rank the schemes and evaluate their merits. To achieve the above objectives, a key issue is to establish a weighted evaluation model for the evaluation indicators. This model should transform qualitative analysis into quantitative values ​​based on the objective importance scaling principle in multi-objective programming to determine the relative weights of the indicators, providing evaluation criteria for each indicator for the fuzzy optimization method.

[0085] The above weight evaluation process requires interaction between decision-makers and the computer. Typically, human involvement necessitates the use of human decision-making experience. Introducing a weight strategy into the decision evaluation model incorporates human decision-making knowledge and experience into the model through mathematical description. This is one of the most significant advantages of scheduling decision evaluation models compared to conventional models that fail to adequately consider expert decision-making experience. The background for considering objective weights is that when comparing multiple objective indicators of a scheme, their importance varies, meaning their impact on the system differs. Therefore, objective weights need to be constructed for their analysis.

[0086] The method for constructing the target weight judgment matrix is ​​as follows: first, establish a hierarchical structure to facilitate decision-makers in determining the weight of each indicator at each level, until the relative weight of each indicator at the bottom level is calculated.

[0087] The following methods are used to determine the weights of the objectives:

[0088] (a) Based on the actual situation of reservoir water level, upstream water inflow, and downstream water inflow, the three targets are artificially ranked in order of importance;

[0089] (b) Analyze and give the importance values ​​U11, U12, and U13 of the first goal (the most important goal) compared to other goals. U1k represents the importance value of the first goal compared to the k-th goal. The first goal is the most important, so its importance value U11 = 0.5 is defined as the first goal compared to the most important goal, and its importance value U13 = 1 is defined as the first goal compared to unimportant goals. Obviously, U1k satisfies:

[0090] 0.5≤U1k≤1

[0091] The value of U12 can be obtained by linear interpolation.

[0092] (c) Determine the importance of each objective:

[0093]

[0094] Normalizing the importance yields the target weight:

[0095]

[0096] A forward reasoning control strategy is adopted: Based on the predicted inflow of water from the "head" reservoir and the measured initial reservoir water level, each rule in the database is matched to identify all available rules, forming a usable rule set (also known as a conflict set). Then, a rule is selected from the rule set using a knowledge ranking method, and the conclusion of this rule is used as the outflow and power generation load of each cascade hydropower station. Based on the water balance condition, the reservoir water level for the next time period is calculated. The updated water level triggers new matches in the knowledge base, thereby determining the water diversion flow and power generation load of each cascade hydropower station for the next time period. This calculation continues until the end of the entire scheduling cycle.

[0097] The following problems may arise when using this scheduling scheme reasoning method:

[0098] 1) Traditional matching methods mostly use full matching, which means that the observed facts must completely match all the premises of the rule before the rule can be activated for reasoning. However, since the daily flow process cannot be exactly the same, it is difficult to find a rule in the rule base using traditional matching methods, that is, it is difficult to find the corresponding optimal solution.

[0099] 2) These rules may conflict. For example, some rules describe the general rules of reservoir operation, while others describe special rules under special circumstances.

[0100] 3) Measurement errors may cause mismatches. The measurement values ​​of inflow and interval flow are inevitably subject to errors, and the calculation of reservoir water level is generally done using approximate formulas, which also contain errors. Therefore, mismatches are very likely to occur, and applying a single rule for reasoning may result in an incorrect solution.

[0101] This invention employs a relaxed matching method, which identifies several rules that best match or are closest to the facts, forming a set of usable rules so that the system can find a correct scheduling implementation scheme.

[0102] The steps for designing an inference algorithm are as follows:

[0103] The usable rule set U1k is constructed as follows:

[0104] Let the premise of the i-th rule in the rule set be: The conclusion is: Let i = {1, 2, …, N}. N is the number of available rules found in the database. Let the measured or predicted inflow, interval flow, and reservoir water level be Q1, S1, and Z, respectively. Then, the distance between the fact and the premise of the rule can be defined by the following formula:

[0105]

[0106] Where (i=1, 2, …, N), if the following conditions are met:

[0107]

[0108] Then this rule is a candidate rule.

[0109] 1) If the number of candidate rules that meet the conditions is less than 5, then all candidate rules are added to the set of available rules;

[0110] 2) If the number of candidate rules that meet the conditions is greater than 5, then the 5 candidate rules with the smallest distance are added to the set of available rules.

[0111] Reasoning and decision-making:

[0112] After obtaining the available rule set based on the relative distance defined in the above formula, the distance between the fact and each rule in the available rule set is recalculated:

[0113] The weight coefficient for each rule is λi, where (i = 1, 2, 3, 4, 5):

[0114]

[0115] Normalization yields:

[0116]

[0117] A weighted voting method is used to derive the reservoir's scheduling strategy. The smaller the distance between the rule's premise and the facts, the larger the weight coefficient of that rule, and the greater the weight of its conclusion in the reasoning; conversely, the greater the distance, the smaller the weight. Therefore, the weight coefficient effectively reflects the differences in importance among the rules. This reasoning method can effectively resolve rule conflicts, and even if there are errors in the calculated values ​​of inflow and reservoir water level, it can ensure that the system can find at least one reasonable scheduling scheme.

Claims

1. A method for recommending intelligent flood control scheduling schemes based on a scheduling business knowledge graph, characterized in that: The method includes: Step 1: Construct a knowledge graph based on scheduling services, including entity recognition, knowledge extraction, and knowledge fusion; Step 2, Hydrological Time Series Anomaly Monitoring: Detecting and correcting outliers in the forecast data; Step 3: Construct a flood control scheduling auxiliary decision-making system based on knowledge graph graph pattern reasoning to find a correct scheduling implementation plan.

2. The method for recommending intelligent flood control scheduling schemes based on scheduling business knowledge graphs according to claim 1, characterized in that: The entity recognition method includes: First, a label set is formulated based on the corpus samples. Each element in the sample sequence is labeled according to the water conservancy features. Based on this, the preprocessed unstructured water conservancy text is labeled to determine the boundaries of words in the text and to determine the classification features for entity recognition. Then, a BiLSTM model is trained. Sentence features are extracted by saving the context information of the whole sentence, and the forward and backward hidden state results in the text are combined. Finally, the output of the BiLSTM model is passed to CRF as input to form a BiLSTM-CRF structure, and the sequence is labeled using context information. In the model training, a method based on word vector combination is used. An automatically acquired dictionary is used to match sentences, and the input character sequence and all latent words are compiled.

3. The method for recommending intelligent flood control scheduling schemes based on scheduling business knowledge graphs according to claim 1, characterized in that: The knowledge extraction methods include: using a combination of pattern matching and co-occurrence network analysis to extract relationships between water conservancy entities. The pattern matching method uses linguistic knowledge to construct entity relationship patterns based on word, part-of-speech, or semantic features of water conservancy entities, and then matches the water conservancy entity text with the patterns to extract relationships. The co-occurrence network analysis method determines the relationships between entities by constructing a co-occurrence matrix. First, it uses statistical methods to extract the frequency of each water conservancy entity in the text, and then analyzes the co-occurrence ratio of different entities in the text. When the co-occurrence ratio of two entities is greater than a certain threshold, it is considered that there is a relationship between the two entities. The relationship between entities is defined based on the type of entity.

4. The method for recommending intelligent flood control scheduling schemes based on scheduling business knowledge graphs according to claim 1, characterized in that: Methods of knowledge fusion include: Entity fusion: First, the similarity between two strings is calculated using the edit distance metric. Then, the similarity of related attributes is analyzed based on the attribute information of the two entities. Finally, the graph structure similarity is calculated based on the graph relationship between the two entities. Through comprehensive analysis of the three similarities, if the similarity exceeds a set threshold, the two entities are fused. Cross-domain fusion of maps: The fusion of water network maps and water conservancy discipline maps includes the fusion of water conservancy principles or laws with water network entities, and the fusion of knowledge between conceptual terms and water network entities; first, entity alignment technology is used to match the local features of entity characters, and then global features are used to match related entities in the two maps.

5. The method for recommending intelligent flood control scheduling schemes based on scheduling business knowledge graphs according to claim 1, characterized in that: The method for detecting outliers is to use an improved K-means clustering algorithm for time series data.

6. The method for recommending intelligent flood control scheduling schemes based on scheduling business knowledge graphs according to claim 1, characterized in that: Constructing flood control and dispatching auxiliary decision-making based on knowledge graph schema reasoning includes: Step 3.1: Knowledge Graph Pattern Reasoning; By extracting patterns from the knowledge graph and performing reasoning, the relationships and patterns hidden in the data are discovered. Step 3.2, Flood Control Scheme Assisted Decision Making: Use inference algorithms to determine the flood control strategy.

7. The method for recommending intelligent flood control scheduling schemes based on scheduling business knowledge graphs according to claim 6, characterized in that: The knowledge graph pattern reasoning method includes: Pattern extraction: Extracting patterns related to queries or knowledge from knowledge graphs. Using graph mining algorithms to find subgraphs that appear frequently in the graph, these subgraphs are considered to represent patterns of a specific knowledge or query. Pattern matching: Finding entities, relationships, and attributes in a knowledge graph that match a query or knowledge, using graph matching algorithms, rule-based algorithms, or graph traversal-based algorithms.

8. The method for recommending intelligent flood control scheduling schemes based on scheduling business knowledge graphs according to claim 6, characterized in that: Methods for determining flood control strategies using inference algorithms include: using a relaxed matching method to find one or more rules that match or are closest to the facts to form a set of available rules, and using a weighted voting method to obtain the reservoir's control strategy.

Images