A method for assisting intelligent decision-making in multi-dimensional safety dispatching of water conservancy and hydropower projects

Abstract

The present invention provides a method for assisting intelligent decision-making of a multi-dimensional safety scheduling scheme for water conservancy and hydropower projects, comprising: step S1: collecting relevant data of a target water conservancy project; step S2: converting text information into Json format to form a generalized fact triple structure for knowledge graph input; step S3: sorting and dividing the operation monitoring images of the water conservancy hub to generate an image representation based on fact triples; step S4: the text data contains the corresponding moment, the image data adds a time mark, or the extracted image is marked according to the time sequence; step S5: using the integration method of text embedding and image embedding to realize the entity relationship embedding based on text image enhancement; step S6: in the text-image integration method, creating an image entity based on text data verification, and performing risk reasoning. Enhance the image encoder's ability to extract risk features of water conservancy project image data and improve the dynamic accuracy of the scheduling scheme.

Description

Technical Field

[0001] The present invention relates to the field of water conservancy and hydropower engineering, and in particular to an intelligent decision-making method for assisting a multi-dimensional safety dispatching scheme of a water conservancy and hydropower engineering project. Background Art

[0002] Water conservancy projects can effectively improve the utilization rate of water resources and promote economic development and ecological environment protection in daily operation and scheduling. However, water conservancy projects are inevitably subject to risks and hidden dangers due to internal or external factors in their daily operation. At the same time, water conservancy projects have functions such as flood control, power generation, shipping, irrigation, and ecology. The cross-disciplinary nature of multiple projects leads to multi-dimensional risk events. Under the background of increasingly complex climate and environmental conditions and heavy scheduling and operation conditions, water conservancy projects use knowledge graph technology and existing safety scheduling solutions to infer the risk transmission path and evolution process, track and predict the risk process from the source of risk, and finally realize the decision-making and selection of safety scheduling plans. Accurate and efficient handling of risks helps maintain the stable operation of the project itself, ensure the safety of life and property of the people in the basin, and promote sustainable economic development.

[0003] As an artificial intelligence model with reasoning ability, the knowledge graph obtains the semantic network of the relationship between entities through the analysis and reasoning of the relationship connection graph between entities. The main design of the knowledge graph modeling process is data extraction and knowledge representation, multi-source heterogeneous knowledge fusion and knowledge reasoning and evaluation. It can organize data and knowledge into a directed graph structure, so that the model has efficient data retrieval capabilities. At the same time, with powerful hardware configuration, the knowledge graph can achieve accurate knowledge reasoning and data intelligence. At present, the knowledge graph has played a key role in the fields of Internet, medical care, e-commerce, and military. Applying the knowledge graph to the risk reasoning of water conservancy projects, so that it can realize machine knowledge expression and reasoning under multidisciplinary and multimodal data, will help improve the level of engineering intelligence.

[0004] At present, the water conservancy project dispatching plan mainly focuses on maximizing benefits, but as some domestic water conservancy projects have been in operation for a long time and large-scale water conservancy projects with complex functions have been put into operation, it is necessary to strengthen the guarantee of safe dispatching of water conservancy projects. The existing risk decision-making plan still needs to rely on the existing procedures, and its flexibility and timeliness may be affected in an emergency, which may lead to instability in the stable operation of the entire water conservancy project.

[0005] In summary, the following shortcomings still exist in the existing technology and improvement measures are proposed: (1) The current risk identification of water conservancy projects mostly adopts methods such as risk event tree, expert experience method, risk probability calculation, etc., there are loopholes in risk chain reasoning, and the research on cross-professional coupled risk transmission is incomplete; (2) The source of past risk data is single and cannot cover large-scale, real-time water conservancy project data. Summary of the invention

[0006] In view of the above problems, the present invention is proposed to provide an intelligent decision-making method for assisting multi-dimensional safety scheduling schemes for water conservancy and hydropower projects that overcomes the above problems or at least partially solves the above problems.

[0007] According to one aspect of the present invention, a method for assisting intelligent decision-making of a multi-dimensional safety scheduling scheme for a water conservancy and hydropower project is provided, the method comprising:

[0008] Step S1: Collect relevant multimodal data of the target water conservancy project;

[0009] Step S2: Convert the text information into Json format to form a generalized fact triple structure for knowledge graph input;

[0010] Step S3: sorting and dividing the water conservancy hub operation monitoring images to generate image representations based on fact triples;

[0011] Step S4: the text data contains the corresponding time, the image data is marked with a time tag, or the extracted image is marked according to the time sequence;

[0012] Step S5: Using the integration of text embedding and image embedding to achieve text-based image enhanced entity relationship embedding;

[0013] Step S6: Using the text-image integration method, create an image entity based on text data verification and perform risk reasoning.

[0014] Optionally, the step S1: collecting relevant data of the target water conservancy project specifically includes:

[0015] Digital and visual information such as operation logs, emergency plans, risk reports, and related dispatching and operation documents.

[0016] Optionally, the step S2: converting the text information into Json format to form a generalized fact triple structure for knowledge graph input specifically includes:

[0017] The risk transmission paths of various events under various working conditions and the corresponding emergency response measures and plans are extracted from the relevant data of water conservancy projects, and the text encoder Text-LSTM-Adam is used to encode the entity descriptive text to obtain the corresponding representation.

[0018] Optionally, the step S3: arranging and dividing the water conservancy hub operation monitoring images to generate image representations based on fact triples specifically includes:

[0019] Organize and divide the water conservancy hub operation monitoring images;

[0020] Use image encoder to model entity image, enhance image features, and use feature extractor combined with matrix projection to obtain corresponding image features;

[0021] The image representation based on fact triples is generated by combining weighted average according to risk specialization.

[0022] Optionally, the step S4: including the corresponding time in the text data, adding a time mark in the image data, or marking the extracted image according to the time sequence specifically includes:

[0023] In order to characterize the superposition of risks in water conservancy hubs at the temporal and spatial levels, the text data contains the corresponding moments, and time annotations should be added to the image data, or the extracted images should be annotated according to the time sequence.

[0024] Optionally, the step S5: using an integrated method of text embedding and image embedding to implement text-based image enhanced entity relationship embedding specifically includes:

[0025] By integrating text embedding and image embedding, we can achieve entity relationship embedding based on text-based image enhancement.

[0026] Natural language supervision signals are used to guide model training to complete the matching between images and texts.

[0027] Optionally, the step S6: creating an image entity based on text data verification and performing risk reasoning in a text-image integration manner specifically includes:

[0028] Under the text-image integration approach, image entities are created based on text data verification;

[0029] By adopting triple embedding based on joint parallel training, the risk knowledge graph of water conservancy projects is obtained and the whole chain of risk transfer is constructed to realize risk reasoning.

[0030] Optionally, encoding the entity descriptive text using the text encoder Text-LSTM-Adam specifically includes:

[0031] The word vector obtained by randomly initializing text a or the word vector obtained based on BERT training is used as input, and the text vector is obtained after network convolution and pooling operations;

[0032] Build a multi-layer LSTM network, use the Densn classifier and softmax all characters.

[0033] The present invention provides an intelligent decision-making method for assisting multi-dimensional safety scheduling schemes of water conservancy and hydropower projects, the decision-making method comprising: step S1: collecting relevant data of target water conservancy projects; step S2: converting text information into Json format to form a generalized fact triple structure for knowledge graph input; step S3: sorting and dividing water conservancy hub operation monitoring images to generate image representations based on fact triples; step S4: the text data contains corresponding moments, and time annotations are added to the image data, or the extracted images are annotated according to the time sequence; step S5: using the integration of text embedding and image embedding to achieve entity relationship embedding based on text image enhancement; step S6: in the text-image integration mode, creating image entities based on text data verification, and performing risk reasoning. A strategy combining moment projection with feature extractor is proposed to enhance the risk feature extraction capability of image encoder for water conservancy project image data, so that different risks can obtain different risk weights under different working conditions to improve the dynamic accuracy of the scheduling scheme.

[0034] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] 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 only 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.

[0036] Figure 1 A structural diagram of the decoder provided in an embodiment of the present invention;

[0037] Figure 2 This is a text prediction structure diagram based on word embedding and cross-layer residual long short-term memory network of the present invention;

[0038] Figure 3 This is a structural diagram of a deep residual shrinkage network based on weight configuration of the present invention;

[0039] Figure 4 This is a structural diagram of the deep residual shrinkage network of the present invention;

[0040] Figure 5 A knowledge graph structure diagram for multimodal data integration of the present invention;

[0041] Figure 6 It is the text entity data of water conservancy hub risk reasoning based on entity-relationship of the present invention;

[0042] Figure 7 A diagram is generated for the water conservancy hub safety dispatch decision support model based on risk reasoning of the present invention. DETAILED DESCRIPTION

[0043] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0044] The terms "comprises" and "having" and any variations thereof in the description embodiments, claims and drawings of the present invention are intended to cover non-exclusive inclusions, for example, including a series of steps or units.

[0045] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0046] Example 1

[0047] like Figure 1 As shown in the figure, a decision-making method for multi-dimensional safety scheduling of water conservancy projects based on risk reasoning

[0048] S1. Collect relevant data of the target water conservancy project, including: operation logs, emergency plans, risk reports, relevant dispatching and operation documents and other digital and image information;

[0049] S2. Convert text information into Json format to form a generalized fact triple structure for knowledge graph input. Extract the risk transmission path of each event under different working conditions and the corresponding emergency response measures and plans from the water conservancy project related data, and use the text encoder Text-LSTM-Adam to encode the entity descriptive text to obtain its corresponding representation;

[0050] S3. Sort and divide the water conservancy hub operation monitoring images, use the image encoder to model the entity image, which requires feature enhancement of the image to improve the feature capture ability of the model, and then use the feature extractor combined with matrix projection to obtain the corresponding image features, and generate an image representation based on fact triples according to risk profession combined with weighted average;

[0051] S4. It should be noted that in order to characterize the superposition of risks in water conservancy hubs at the temporal and spatial levels, the text data should contain the corresponding time, and the image data should be marked with time, or the extracted images should be marked according to the time sequence;

[0052] S5. In order to improve the specific cognition of risk events by water conservancy hub staff, the present invention adopts the integration of text embedding and image embedding to realize entity relationship embedding based on text-based image enhancement. Contrastive Language-Image Pre-training (CLIP) is introduced for image-text matching, that is, natural language supervision signals are used to guide model training to complete the matching between image and text;

[0053] S6. Create image entities based on text data verification by adopting the text-image integration method. Finally, the risk knowledge graph of water conservancy projects is obtained and the whole chain of risk transmission is constructed to realize risk reasoning by using triple embedding based on joint parallel training.

[0054] The specific implementation strategies in the above steps are as follows:

[0055] In S2, the optimization steps of running the Text-LSTM-Adam text encoder are as follows:

[0056] A. Take the word vector obtained by randomly initializing text A or the word vector obtained based on BERT training as input, and obtain the text vector after network convolution, pooling and other operations. The structure of the text encoder based on time series features constructed by the present invention is as follows Figure 1 shown.

[0057] Among them, manual verification and annotation are used to obtain entity text descriptions from the annotation files of the risk text image corpus, and the termination words in the entity text are deleted. This process uses LSTM and word embedding to train the network and predict the next character. The word prediction structure diagram with word embedding function and cross-layer residual long short-term memory network is shown in Figure 2 .

[0058] LSTM is more sensitive to time series data and more efficient than RNN. The calculation equations of the forget gate, input gate, and output gate of its model are:

[0059] f t =σ(W f ·[h t-1 ,x t ]+b f ) (1)

[0060] i t =σ(W i ·[h t-1 ,x t ]+b i ) (2)

[0061] o t =σ(W o ·[ht-1 ,x t ]+b o ) (3)

[0062]

[0063]

[0064]

[0065] In the formula, f t 、i t and t Respectively represent the calculation results of the forget gate, input gate and output gate at time t; W f , W i and W o Represents the weight matrices of the forget gate, input gate, and output gate respectively; b f , b i and b o Represent the bias of the forget gate, input gate and output gate respectively; σ represents the activation function sigmoid; [h t-1 ,x t ] represents the state h of the hidden layer at time t-1 t-1 and the sequence input x at time t t Two vectors are concatenated into a longer vector, c t represents the state at time t; h t represents the state of the hidden layer at time t; x t represents the current sequence input at time t; c t-1 represents the state of the memory layer at the previous moment t-1, h t-1 Represents the state of the hidden layer at the previous time t-1, and the storage module outputs the storage unit state and the hidden layer state at time t. represents the state of the input unit at time t; W c represents the input unit state weight matrix; b c Represents the bias of the input state; tanh represents the activation function ☉, i t represents the calculation result of the input gate at time t, o t Represents the calculation result of the output gate at time t.

[0066] The text a is preprocessed using the Python programming language, where Table 1 provides pseudo code for text generation with a character length of 120 and a sampling sequence of 3 characters. Sequences of length maxlen are extracted, one-hot encoded, and packaged into a three-dimensional Numpy array with a shape of (sequences, maxlen, unique_characters).

[0067] Table 1. Character extraction and text generation based on One-hot encoding

[0068]

[0069] b. Build a multi-layer LSTM network, use the Densn classifier and perform softmax on all characters. After one-hot encoding, the model uses categorical_crossentropy as the loss evaluation function. The Adam optimizer is used to optimize the model in this process so that the training set can be used to evaluate the model after each period.

[0070] The number of hidden units in the LSTM hidden layer is defined as 256, the input is a single feature (one character corresponds to one integer), and the probability of the dropout layer in the model is set to 0.2. The LSTM output is a unit, whose first element is the hidden state of the LSTM at each time step. The last hidden layer will contain the most information and be passed to the output layer, which is a fully connected layer to generate the logic of the 120 vocabulary (this process can be converted into a probability prediction using softmax).

[0071] In S3, a deep residual shrinkage network based on cross entropy loss weights is used to construct entity image feature representation to obtain entity image-based representation. Specifically, the image encoder is used to construct entity image features. Figure 3 The structural framework of the image encoder is given.

[0072] The framework uses a deep residual shrinkage network based on cross entropy loss weights as a feature extractor for entity images, then uses the orthogonal projection matrix method to obtain entity-based image features and calculates the weighted average of the image based on the target risk. The specific steps are:

[0073] a. Collect monitoring images of various disciplines and departments of the water conservancy hub, and arrange and classify the images based on time sequence. If some images are missing, the corresponding images can be obtained by frame-by-frame extraction of monitoring videos. In this part, the state images (image1, image2, image3...) of a certain entity of the water conservancy hub in multiple time sequences will be used as encoder input;

[0074] b. Use the deep residual shrinkage network to perform feature enhancement and feature extraction on the image. The schematic diagram of the deep residual shrinkage network is shown in Figure 4 The cross-layer connection in the Deep Residual Shrinkage Network (DRSN) can realize the forward transmission and reverse traceback of data features, which can improve the training efficiency of the model and avoid problems such as gradient explosion. Its structure is as follows Figure 1As shown in the figure, the data is integrated after dimension reduction in the input layer and then input into the convolution layer. The convolution kernel extracts useful feature information and passes it to the residual contraction layer. Each feature map has one or more feature channels to satisfy the network's feature extraction of multiple attention points in the image and avoid feature loss. At the same time, the gap layer of the network can reduce the number of weights in the connection layer to maintain network stability.

[0075] In the weight configuration stage, the present invention constructs a weight configuration strategy based on multiple attention targets to enhance the network's weight allocation to different risk points in the image when facing different risks, thereby achieving high-precision risk feature extraction. The specific steps are as follows:

[0076] (1) The multi-objective weight configuration strategy is to perform pairwise comparison when identifying risk features in an image. Its mathematical expression is:

[0077]

[0078] Among them, w i,j Representing image features and The comparison results are represents the importance of target i, Indicates the importance of target j. In the risk reasoning and disposal stage, the staff of the water conservancy hub needs to quickly obtain the highest level risk characteristics and make corresponding disposals. However, some risks are of low importance in the early initial stage and may be identified as 0. In order to prevent the risk from being ignored and causing misjudgment, the present invention sets the potential low risk as w n+1 .

[0079] (2) When comparing two risk characteristics, the risk characteristics with greater severity, deterioration trend and impact consequences will be limitedly extracted, so their weights should be greater than other risks. Their comparison values ​​can be expressed as:

[0080] w i,n+1 =1,i=1,2,3,…,n (8)

[0081] Then the weight value that can be obtained for any target risk feature is expressed as:

[0082]

[0083] In the formula, w i Represents the weight obtained by the target risk feature. The weight of the most dangerous risk feature can be expressed as:

[0084]

[0085] c. Use the shared orthogonal projection matrix method to perform entity projection on the acquired features to obtain the features of the entity corresponding to the image. The features of the i-th entity image can be expressed as:

[0086] p i =M·f(image i ) (11)

[0087] Among them, f represents the convolution part in the deep residual contraction network, and M represents the orthogonal projection matrix used for pre-training.

[0088] S6. Text / Image-Entity / Relationship Modeling

[0089] The generated text and image representations are integrated into knowledge graph learning, and entity-relationship modeling is completed using transfer learning methods and translating embedding (TransE).

[0090] a. Construct a modeling scoring function, the formula is as follows:

[0091] F(h, r, t) = E s +αE d +βE i (12)

[0092] In the formula, α and β are hyper parameters, E s is the evaluation function of the structure, E d is defined as follows:

[0093] E d =||h s +rt d ||+||h d +rt s ||+||h d +rt d || (13)

[0094] In the formula, the first and second terms are the text representations in the head entity and the tail entity, the other one is represented in structural form, and the third term h and t are both text-based representations. The evaluation function of the image is the same as above, using shared relationship embedding, and projecting the orthogonal matrix of the entity representation based on structure, text, and image into a continuous low-dimensional vector space, and making them associated with each other through transfer learning and joint learning.

[0095] b. Set the model parameters, X represents words, E represents entities, R represents relation embedding, M represents the orthogonal projection matrix in the image encoder, X is obtained through random initialization, Json processing, and BERT pre-training, and E and R are obtained through deep residual contraction network and LSTM network pre-training.

[0096] c. To improve the representation of learning entities, relations, texts, and images to achieve parameter optimization, a loss function based on the minimum margin is used for target training:

[0097]

[0098] Where γ represents the interval hyperparameter, [x] + To keep the integer part of x, S is the correct triple, and S' is the set of incorrect triples of S. S' is constructed by replacing the correct triple (h, r, t) with a probability-based entity to obtain (h', r, t'), ω and is the error weight hyperparameter, The purpose of this loss function is to make the correct triplet F(h, r, t) as small as possible and the incorrect triplet as large as possible so that the overall error loss approaches 0.

[0099] The knowledge graph structure diagram of multimodal data integration, such as Figure 5 shown.

[0100] Example 2

[0101] S1. Set data sequence, collect and organize parameters such as power generation, voltage, rainfall, water level, power generation flow, discharge flow, unit swing, unit amplitude, volute pressure, tailwater pipe pressure, generator three-phase voltage, main transformer three-phase voltage, generator winding temperature, emergency plan, operation log, insulator flashover image, transformer infrared temperature measurement image, etc. of hydraulic turbines in water conservancy projects;

[0102] S2. Embed entity text and entity images with additional information into the knowledge graph to construct a multimodal knowledge graph dataset. At the image level, the electromechanical equipment status monitoring images in the water conservancy hub are constructed in a form similar to WN9IMG; at the text level, the methods of S2 and S4 in the invention content are used to generate corresponding text-relation entities. The text-based relationship-entity formed by the present invention is as follows: Figure 6 shown.

[0103] S3. Taking TransE and TransR as the baseline, the entity-machine relationship is embedded in the dimension, and the word embedding dimension and interval hyperparameter γ are set.

[0104] S4. In the representation layer, the deep residual shrinkage network with weighted cross entropy function in the invention is used to extract the key area features of the image, and the maximum area correlation value and the single image feature dimension value are set. Then, the text vector is mapped to the single image feature dimension value using LSTM. Then, the number of iterations and batch size are set, and the optimizer is selected as Adam.

[0105] S5. In the multimodal attention layer, adjust the number of encoder and decoder layers and the head value and latitude value in the multi-head attention mechanism so that the image and text are consistent.

[0106] S6. Use the evaluation criteria in the visual challenge to preliminarily evaluate the accuracy of the model, and use the water conservancy hub risk log and water conservancy hub staff and experts to further verify the model results.

[0107]

[0108] The following is an application study of a large water conservancy hub in the Yangtze River Basin. The model hardware configuration is Intel I9-13900k, NVIDIA RTX 4090, 32G memory, and runs in python3.9 environment. The implementation background of this case is the professional risk reasoning and emergency plan comparison and decision-making of water conservancy hub flood control, dam safety, electromechanical, etc. under the condition of medium-term (5-7 days) continuous heavy rainfall. Figure 7 The risk map of water conservancy hubs based on risk reasoning is displayed. Clicking on any node will obtain the corresponding risk transmission path, management department, emergency measures and risk plan. In the context of this case, the risk reasoning diagram gives key inspection targets and observation indicators, among which flood discharge and partial discharge of transformer insulators become potential high-risk risks, and secondary potential risks are hydro-generators, plant power systems, main connections for electrical transmission, and downstream water level fluctuations, which trigger the cascade flood control emergency dispatch plan, electromechanical equipment failure plan, and full-site power outage emergency plan.

[0109] Beneficial effects: The knowledge graph can clarify the risk causes of water conservancy projects, key nodes in the risk transmission chain, expected paths of risk transmission, and emergency response measures and plans for risks under the corresponding risk transmission paths. The risk development path of the hydropower station and the risk evolution trend after intervention measures are determined based on the key nodes in the risk transmission chain, monitoring information, emergency response measures and risk plans. The combination of text and images can enable dispatching and operation personnel to obtain more specific risk status.

[0110] The use of data-driven and knowledge graphs to construct a dynamic evolution model of water conservancy project risks makes up for the shortcomings of the existing water conservancy project risk reasoning chain. Even with a small amount of data, a knowledge graph of water conservancy hub risk reasoning can be constructed. And risk reasoning based on data-driven avoids the shortcomings of lack of objectivity associated with the risk inference process relying solely on expert experience. Using this model, water conservancy hub staff can achieve cross-professional risk prevention and disposal, and the risk transmission chain covers a wide range, which can make up for the lack of experience of dispatchers in dealing with unfamiliar professions and sudden rare risks, and effectively improve the efficiency and intelligence of water conservancy hub risk emergency plan decision-making.

[0111] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for assisting intelligent decision-making in multi-dimensional safety scheduling of water conservancy and hydropower projects, characterized in that: The decision-making method includes: Step S1: Collect relevant multimodal data of the target water conservancy project; Step S2: Convert structured / unstructured text data into Json format to form a generalized fact triple structure for knowledge graph input, including: The risk transmission paths of various events under various working conditions and the corresponding emergency response measures and plans are extracted from the relevant data of water conservancy projects, and the text encoder Text-LSTM-Adam is used to encode the entity descriptive text to obtain the corresponding representation; Step S3: Arrange and divide the water conservancy hub operation monitoring images to generate image representations of fact triples based on time series or causal relationships, including: Organize and divide the water conservancy hub operation monitoring images; Use image encoder to model entity image, enhance image features, and use feature extractor combined with matrix projection to obtain corresponding image features; Generate an image representation based on fact triples by combining weighted average according to risk specialties; Step S4: the text data contains the corresponding time, the image data is marked with a time tag, or the extracted image is marked according to the time sequence; Step S5: Using the integration of text embedding and image embedding to achieve text-based image enhanced entity relationship embedding; Step S6: Using the text-image integration method, create an image entity based on text data verification and perform risk reasoning.

2. The method for assisting intelligent decision-making of a multi-dimensional safety scheduling scheme for water conservancy and hydropower projects according to claim 1 is characterized in that: The step S1: collecting relevant data of the target water conservancy project specifically includes: Digital and image information of monitoring data, operation logs, emergency plans, risk reports, and related dispatching and operation documents.

3. The method for assisting intelligent decision-making of a multi-dimensional safety scheduling scheme for water conservancy and hydropower projects according to claim 1 is characterized in that: The step S4: including the corresponding time in the text data, adding time annotation to the image data, or annotating the extracted image according to the time sequence specifically includes: In order to characterize the superposition of risks in water conservancy hubs at the temporal and spatial levels, the text data contains the corresponding moments, and time annotations should be added to the image data, or the extracted images should be annotated according to the time sequence.

4. The method for assisting intelligent decision-making of a multi-dimensional safety scheduling scheme for water conservancy and hydropower projects according to claim 1 is characterized in that: The step S5: using the integration of text embedding and image embedding to implement text-based image enhanced entity relationship embedding specifically includes: By integrating text embedding and image embedding, we can achieve entity relationship embedding based on text-based image enhancement. Natural language supervision signals are used to guide model training to complete the matching between images and texts.

5. The method for assisting intelligent decision-making of a multi-dimensional safety scheduling scheme for water conservancy and hydropower projects according to claim 1 is characterized in that: The step S6: creating an image entity based on text data verification and performing risk reasoning in a text-image integration manner specifically includes: Under the text-image integration approach, image entities are created based on text data verification; By adopting triple embedding based on joint parallel training, the risk knowledge graph of water conservancy projects is obtained and the whole chain of risk transfer is constructed to realize risk reasoning.

6. The method for assisting intelligent decision-making of a multi-dimensional safety scheduling scheme for water conservancy and hydropower projects according to claim 1 is characterized in that: The encoding of entity descriptive text by using the text encoder Text-LSTM-Adam specifically includes: The word vector obtained by randomly initializing text a or the word vector obtained based on BERT training is used as input, and the text vector is obtained after network convolution and pooling operations; Build a multi-layer LSTM network, use the Densn classifier and softmax all characters.

Images