CN-121998296-A - Irrigation area water distribution method based on topology and hierarchy perception graph similarity calculation

CN121998296ACN 121998296 ACN121998296 ACN 121998296ACN-121998296-A

Abstract

The application provides a pouring area water distribution method based on topology and hierarchy perception graph similarity calculation, which relates to the field of intelligent water distribution, and comprises the steps of respectively modeling a historical water distribution scheme and a current water demand graph into directed non-weighted graphs; the method comprises the steps of carrying out multidimensional feature coding on each node, obtaining node enhancement representation after fusion, carrying out multi-scale feature extraction on the node enhancement representation by utilizing a relational graph convolution network, introducing a gating mechanism to filter invalid node information to obtain final node embedding, carrying out weighted attention aggregation on dual-path features by adopting hierarchical constraint based on the final node embedding to generate graph level representation, calculating similarity scores between two graphs, carrying out weighted fusion to obtain a predictive graph editing distance, and recommending an optimal water distribution scheme from a historical scheme library according to the predictive graph editing distance. According to the technical scheme, the matching task of the historical water distribution scheme and the current demand diagram is converted into the scene adaptation type diagram similarity modeling problem, and a core basis is provided for intelligent recommendation of the water distribution scheme of the irrigation area.

Inventors

LI XINCHUAN
LU CHENGHAO
YAO HONG
DONG LIJUN
LIANG QINGZHONG
LI SHENGWEN

Assignees

中国地质大学（武汉）

Dates

Publication Date: 20260508
Application Date: 20251222

Claims (8)

1. The irrigation area water distribution method based on topology and hierarchy perception graph similarity calculation is characterized by comprising the following steps of: The method comprises the following steps of S1, respectively modeling a historical water distribution scheme and a current water demand graph into a directed non-weighted graph, wherein nodes in the directed non-weighted graph represent main channels or branch channels, edges represent water flow directions, and node attributes comprise channel system length, farmland area, design flow and water demand; S2, carrying out multidimensional feature coding on each node, including hierarchical coding, attribute coding and random walk position coding, and obtaining node enhancement representation after fusion; S3, performing multi-scale feature extraction on the node enhancement representation by using a relational graph convolution network, and introducing a gating mechanism to filter invalid node information to obtain final node embedding; s4, based on final embedding of the nodes, adopting hierarchical constraint weighting attention aggregation dual-path characteristics to generate graph-level representation; S5, calculating similarity scores between the two graphs through graph level representation and final node embedding, and obtaining the editing distance of the predictive graph through weighted fusion; and S6, recommending an optimal water distribution scheme from the historical scheme library according to the edit distance of the predictive diagram.
2. A method of irrigated area water distribution based on topology and hierarchical perceptron similarity calculation as recited in claim 1, wherein step S1 comprises: Input device And Respectively representing the historical water distribution scheme of a single irrigation area and the directed non-weighted graph of the current water demand graph in the knowledge base, Wherein: Is that Or (b) ; For a set of nodes, Representing the nodes of the main channel, Representing a branch canal node; Is that Or (b) , The water flows only from the main channel to the branch channel; Is that Or (b) , Is the adjacent matrix of (a) If it meets Then Otherwise ; Definition of the degree matrix Is a diagonal matrix with diagonal elements , Is a node And meet the following 、 ; For the following Is provided with To ensure that Is the reversibility of (2); node validity markers in a directed non-weighted graph defining a historical water distribution scheme , Corresponding to Or (b) Is provided with a node which is connected with a node, Indicating that the node is active, i.e. water demand And design flow rate ; Indicating that the node is not enabled and the parameter is not valid.
3. A method of irrigated area water distribution based on topology and hierarchical perceptron similarity calculation as claimed in claim 2, wherein step S2 comprises: The random walk position encoding includes: construction of random walk matrix Wherein In order to be a contiguous matrix, Is an output matrix; From the slave Step random walk matrix Extracting node self-access probability from diagonal elements as initial position characteristics; and mapping the initial position features into dense vectors through a multi-layer perceptron, and splicing the dense vectors with the node attribute features to obtain the node enhancement representation.
4. A method of irrigated area water distribution based on topology and hierarchical perceptron similarity calculation as claimed in claim 3, wherein step S3 comprises: uniquely identifying the side relationship of the directed non-rights graph of the historical water distribution scheme as Wherein Representing a set of main trench nodes , Representing a set of canal nodes ; Assigning a learnable weight matrix to a side relationship Dimension and node initial embedding Keeping consistency; taking into account the dominant position of the main channel in water distribution decision, introducing a hierarchical weighting factor : When the main channel nodes aggregate neighbor information, the weight matrix is as follows ; The branch channel node has no edge neighbor, and the weight matrix inherits by default ; The relation graph rolling network is The aggregation formula of the layers is: Wherein, the , The function is activated and the function is activated, Is the first The self-loop weight matrix of the layer is used for reserving characteristic information of the node itself and avoiding original attribute loss in the aggregation process; Is a node In relation to A lower neighbor set; Is a normalized coefficient; representing the Hadamard product; A learnable weight matrix that is a side relationship; representing a residual connection; activating a function for Sigmoid; Is a gating factor; embedding an initial node And splicing the output characteristics of the layer relation graph convolution network to form a node to be finally embedded.
5. A method of irrigated area water distribution based on topology and hierarchical perceptron similarity calculation as recited in claim 4, the method is characterized in that the step S4 comprises the following steps: Computing node final embedded mean vector as global context feature : Wherein, the As a set of nodes of the graph, As a total number of nodes, Is a node Is finally embedded in the node of (c) a, For the dimension of the node characteristics, Reflecting the overall attribute distribution and topology association characteristics of the full irrigation area; computing nodes considering fusion of global semantics and canal system priority Is given by: Wherein, the Mapping node characteristics and global context to the same characteristic space for a learnable weight matrix to enhance the rationality of association degree calculation; Scoring vectors for attention for quantifying the strength of association of node features with global context; for identifying canal system level, main canal Canal ; For the level enhancement coefficients, ensuring that the main channel occupies dominant weight in the graph level representation by amplifying the score value of the main channel node; Is a nonlinear activation function; converting the attention score into a normalized weight by a softmax function; and carrying out weighted summation on all node characteristics based on the normalized attention weight to obtain a graph level representation of the focused core node semantics, wherein the formula is as follows: And introducing an unbiased node characteristic sum aggregation strategy, wherein the formula is as follows: Wherein the method comprises the steps of No additional parameters need to be learned; dynamic fusion of dual path features by dimension-by-dimension learnable weight vectors to obtain final graph level representation The following are provided: Wherein, the For dimension-by-dimension weight vector learning, the element value range is Support the dynamic change adjustment of model according to the water distribution scene of the irrigation district And (3) with Is a fusion ratio of (2); for Hadamard product, the feature weight of each dimension in the fusion process is ensured to be independently adjustable, and finally generated The method has core water distribution semantics and global scale characteristics and is used for subsequent similarity calculation.
6. A method of irrigated area water distribution based on topology and hierarchical perceptron similarity calculation as recited in claim 5, the method is characterized in that the step S5 comprises the following steps: The node level similarity score is calculated as follows: The attribute matching degree and the local topology association strength of any two nodes are quantized by adopting cosine similarity, and the two nodes are respectively from And Meanwhile, invalid node noise is filtered through a gating mechanism, and the specific formula is as follows: Wherein, the For a vector dot product embedded by two nodes, 、 Respectively embedding L2 norms for two nodes, wherein the range of the calculated result value is ; In order to indicate the function, Representing logical and operations, only when both nodes are active Otherwise ; Based on the filtered similarity value Constructing a pairwise node interaction matrix , 、 Respectively are 、 Node total, matrix element The fine granularity matching information of the effective nodes is completely reserved, and the dimension is consistent with the total number of the nodes of the two graphs; Extracting node interaction matrix The local strong correlation feature and the global distribution feature: pair interaction matrix Execution of Window maximum pooling operation, setting the sliding step length to be 1, and reserving the maximum similarity value in each local window to obtain a pooling matrix The strong association characteristic among the nodes is enhanced, and weak association interference is restrained; pooling matrix Respectively calculating a row average value and a column average value to obtain a row characteristic vector And column feature vector , wherein, Reflecting Middle (f) Personal node The average matching strength of all the nodes is, Reflecting Middle (f) Personal node Average matching strength of all nodes; Line feature vector And column feature vector Splicing, inputting the fusion of nonlinear characteristics to a multi-layer perceptron MLP, and mapping the fusion into a scalar, wherein the formula is as follows: Wherein, the The layer weight matrix is hidden for the MLP, Bias items for hidden layers; for the MLP output layer weight matrix, Realizing nonlinear transformation for activation function and final output The node level score is in a scalar form, the numerical value is inversely related to the matching degree of the node attributes and the local topology of the two graphs, and the lower the score is, the higher the matching degree is; The similarity score at the graph level is calculated as follows: The water distribution scheme diagram of the two irrigation areas to be matched is set as 、 Their corresponding graphic level representations are respectively 、 Capturing complex semantic association between graph-level representations by fusing tensor second-order interaction and linear interaction, wherein the formula is as follows: Wherein, the Representing traversal tensors A kind of electronic device Two-dimensional slice , Calculation of Second order interaction values under each slice, forming A dimension tensor interaction feature; is a linear interaction weight matrix; Is that And (3) with Is a concatenation vector of (a); The high-order interaction characteristics are fused; Representing the bias vector; will be by MLP The dimensional interaction feature maps to a similarity score, the formula: Wherein, the In order to score the vector of the value, Is biased; Similarity score for the graph level; scoring similarity at node level Similarity score to level of graph The weight weighting fusion can be learned to obtain And (3) with Predictive map edit distance of (2) The formula is: Wherein, the Is a learnable weight; To finally predict GED value, value size and value And (3) with Is inversely related to the overall degree of matching.
7. An electronic device comprising a processor, a memory, a user interface and a network interface, the memory for storing instructions, the user interface and the network interface for communicating to other devices, the processor for executing the instructions stored in the memory to cause the electronic device to perform the method of irrigated area water based on topology and hierarchy awareness graph similarity calculations of any one of claims 1-6.
8. A computer readable storage medium storing instructions which, when executed by a computer, perform the method of irrigated area water distribution based on topology and hierarchical perceptron similarity calculation of any one of claims 1-6.

Description

Irrigation area water distribution method based on topology and hierarchy perception graph similarity calculation Technical Field The application relates to the field of intelligent water distribution, in particular to a pouring area water distribution method based on topology and hierarchy perception graph similarity calculation. Background In the water resource allocation of the irrigation areas, the scientific recommendation of the water distribution scheme is a key link for improving the water resource utilization efficiency and guaranteeing the agricultural production. For a long time, the main stream thought for making a water distribution scheme in a irrigated area is a multi-objective optimization method, and the optimal water distribution scheme is solved by constructing a plurality of objective functions such as water demand satisfaction, water distribution cost, water resource utilization rate and the like and combining the conditions such as canal flow constraint, irrigation time constraint and the like. On the one hand, the method needs to set the weight coefficient of each target manually and subjectively, and the weight adjustment difficulty of different crops and different growth periods is high, so that the scheme suitability is obviously influenced by human factors, and on the other hand, the method is more critical in that the method does not effectively utilize the existing successful water distribution scheme, can not reuse the history engineering experience, but solves the numerical optimal solution from zero each time, thereby omitting the canal system adaptation logic contained in the history scheme, and also causing low decision efficiency, and being difficult to meet the real-time water distribution requirement of the irrigation area. In the application scene of the research, the core complaint of the water distribution scheme is that the scheme matched with the current scene is quickly matched based on historical successful experience. The topology relation of canal systems, water demand distribution of each canal system, water flow distribution proportion and the like in the water distribution scheme of the irrigation area are related problems of entity and relation, and the graph structure can naturally restore the core characteristics, particularly the directionality of water flow flowing from a main canal to a branch canal, and is a core carrier of water distribution logic. The graph similarity calculation method can realize the efficient multiplexing of the historical experience by comparing the graph structural similarity of the historical scheme in the knowledge base and the current scene, and compared with the thought of solving from zero by multi-objective optimization, the graph similarity calculation method realizes the multiplexing of the historical experience. However, the existing graph similarity calculation method is not adapted to the core characteristics of the irrigation area scene and is difficult to directly apply. In traditional graph similarity index calculation, graph Editing Distance (GED) is a common method, similarity is measured by calculating the minimum number of modification operations between two graphs, but the method belongs to the NP-hard problem, when the node level in the graph is increased, the calculation complexity is exponentially increased, and the real-time matching requirement cannot be met. Although pruning strategies and heuristics have been used for approximation calculations, the uncertainty of sub-optimal solutions still limits their application in engineering scenarios. With the development of deep learning technology, a method based on a graph neural network is generated, and although the method improves the calculation efficiency through feature extraction, the core pain point of the irrigation area scene is not solved pertinently. The method is characterized in that SimGNN is used as an early representative model, position information of nodes in channel system topology is not effectively captured, water distribution proportion can be directly influenced by branch channels at different positions at the downstream of a main channel, meanwhile, edge orientation is ignored, the directional logic of water flow in an irrigation area is the core of a water distribution scheme, characteristic characterization is incomplete due to the fact that the two are missing, matching precision is insufficient, an ERIC framework is provided in recent years, training efficiency and judging capability are optimized through aligning a regularization module and a multi-scale GED (get) judging device, but hierarchical characteristics of the channel system in the irrigation area are not adapted, functional differences and water flow direction constraints of the main channel and the branch channels cannot be accurately captured, single-dimension fusion is relied on in a similarity calculation link, cooperative influence of node attributes, topological po