Artificial intelligence-based hydroelectric project management system and method

Abstract

The present application relates to the technical field of hydropower project management, in particular to a hydropower project management system and method based on artificial intelligence, comprising: collecting real-time monitoring data of each construction area of the hydropower project to construct an original time sequence feature set; inputting the original time sequence feature set into a graph neural network to mine spatial dependency relationships of each construction area and generate a regional correlation feature map; performing time sequence convolution operation on the regional correlation feature map to capture construction state change rules under multiple time windows and construct a dynamic state tensor; introducing a multi-head attention mechanism to calculate risk weight distribution in the decision cycle of each construction area and output a weighted risk vector; and comparing the weighted risk vector with a preset construction progress benchmark matrix to automatically generate resource scheduling correction instructions adapted to the on-site working conditions, thereby realizing intelligent risk assessment and dynamic resource allocation of the hydropower project.

Description

Technical Field

[0001] This invention relates to the field of hydropower project management technology, and in particular to a hydropower project management system and method based on artificial intelligence. Background Technology

[0002] Hydropower projects cover a wide area and involve numerous work zones, with inherent connections between these zones through processes and resource sharing. Currently, hydropower project management largely relies on manual analysis combined with single-time-series data analysis, processing monitoring data for individual work zones without conducting comprehensive multi-zone correlation analysis. This approach, where various monitoring time-series data are recorded and processed independently, fails to uncover the hidden spatial dependencies between different work zones and cannot generate a characteristic representation of the structural connections between these zones.

[0003] Existing data analysis methods are mostly limited to statistical analysis of numerical changes within fixed time intervals, lacking the ability to capture the continuous evolution of construction status across multiple time windows. They also struggle to simultaneously integrate spatial structural information and temporal variation information to construct a unified state data tensor. Furthermore, project risk assessment consistently employs a uniform standard, making it impossible to adaptively distribute risk weights across different areas based on real-time construction conditions.

[0004] Traditional resource scheduling relies on fixed schedule standards and manual experience, lacking an automated process for comparing the quantitative risk vector with the schedule baseline matrix. It also cannot automatically generate corresponding scheduling adjustment instructions based on real-time on-site conditions. Traditional management models lack spatial correlation analysis, temporal dynamic modeling, and adaptive risk allocation capabilities, resulting in highly subjective control decisions. This fails to meet the application requirements of large-scale hydropower projects, including multi-regional collaborative construction, dynamic risk identification, and automated, precise resource allocation. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an artificial intelligence-based hydropower project management system and method.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an artificial intelligence-based hydropower project management method, comprising:

[0007] Collect real-time monitoring data from various construction areas of the hydropower project and construct an original time-series feature set;

[0008] The original temporal feature set is input into a graph neural network to extract the spatial dependencies between construction areas and generate a regional association feature map.

[0009] Temporal convolution operations are performed on the region association feature map to capture the state change patterns of each construction area under different time windows and form a dynamic state tensor.

[0010] The dynamic state tensor is passed to the multi-head attention mechanism to calculate the risk weight distribution of each construction area in the current decision-making cycle and output the weighted risk vector.

[0011] The weighted risk vector is compared with the preset construction progress benchmark matrix to generate a resource scheduling correction instruction.

[0012] As a further aspect of the present invention, the step of collecting real-time monitoring data from various construction areas of a hydropower project and constructing an original time-series feature set specifically includes:

[0013] Multiple sensor nodes are deployed in each construction area of ​​the hydropower project to synchronously collect environmental parameter data and equipment operating status data;

[0014] Add a unified timestamp to the environmental parameter data and equipment operating status data collected by each sensor node, and then transmit the data with the unified timestamp to the edge computing gateway;

[0015] At the edge computing gateway, data from different sensor nodes are grouped and aggregated according to the construction area number;

[0016] For each grouped aggregated data, a missing value detection operation is performed, and missing values ​​are filled using the mean of data from the same construction area at adjacent timestamps.

[0017] After filling in the missing values, the data for each construction area are arranged into a time series in ascending order of timestamps.

[0018] The time series sequences of all construction areas are aligned according to the construction area number and construction area name and combined into a multidimensional array as the original time series feature set.

[0019] As a further aspect of the present invention, the missing value detection operation specifically includes using the isolated forest algorithm to identify abnormal missing values ​​caused by sensor malfunctions.

[0020] As a further aspect of the present invention, the step of inputting the original temporal feature set into a graph neural network to extract the spatial dependencies between construction areas and generate a regional association feature map specifically includes:

[0021] Each construction area is taken as a graph node, and the time sequence of each construction area in the original time sequence feature set is taken as the initial feature vector of the graph node;

[0022] Based on the overall construction layout plan of the hydropower project, determine the physical connection relationship between each node in the plan and construct an undirected edge between two nodes that have a physical connection relationship.

[0023] Combine all graph nodes and all undirected edges into an initial graph structure;

[0024] The initial graph structure and the initial feature vector of each graph node are input into the message passing layer of the graph neural network;

[0025] In the message-passing layer of the graph neural network, the feature vectors of the neighboring graph nodes of each graph node are aggregated to update the feature representation of that graph node;

[0026] Repeatedly executing message passing operations allows feature information to be fully diffused among nodes in the global graph.

[0027] After message passing is completed, the updated feature vectors corresponding to each graph node are arranged into a two-dimensional matrix according to the graph node order, which serves as the region association feature map.

[0028] As a further aspect of the present invention, the graph neural network is a graph attention network.

[0029] As a further aspect of the present invention, the step of performing a temporal convolution operation on the region association feature map to capture the state change patterns of each construction area under different time windows and form a dynamic state tensor specifically includes:

[0030] The region association feature map is divided into multiple continuous and non-overlapping fixed-length time windows along the time dimension;

[0031] Configure an independent temporal convolution kernel for each fixed-length time window and set the size of each temporal convolution kernel to be equal to the length of the time window;

[0032] The convolutional output features of that time window are obtained by multiplying each temporal convolution kernel element-wise with the region-related feature map within its corresponding fixed-length time window and then summing the results.

[0033] The convolutional output features of all time windows are reassembled into a three-dimensional feature tensor in chronological order.

[0034] Batch normalization is performed on the three-dimensional feature tensor to eliminate dimensional differences between different construction areas;

[0035] The three-dimensional feature tensor after batch normalization is used as the dynamic state tensor.

[0036] As a further aspect of the present invention, the step of passing the dynamic state tensor to a multi-head attention mechanism to calculate the risk weight distribution of each construction area within the current decision-making cycle and outputting a weighted risk vector specifically includes:

[0037] The dynamic state tensor is split into multiple independent region state vectors along the construction region dimension.

[0038] In each attention head, a linear transformation is performed on the state vector of each region to generate the query vector, key vector, and value vector for that region.

[0039] In each attention head, the similarity score between the query vector of each region and the key vector of all regions is calculated, and the similarity score is converted into attention weights using a normalized exponential function;

[0040] In each attention head, the value vectors of all regions are summed according to their corresponding attention weights to generate the risk-weighted vector for that attention head;

[0041] After concatenating the risk weighted vectors generated by all attention heads along the feature dimension, a linear transformation is performed to generate a fused risk weighted vector;

[0042] Each element in the fused risk weighted vector is used as the risk weight of the corresponding construction area in the current decision-making cycle, and the risk weights of all construction areas are arranged in the order of the area index to form the weighted risk vector.

[0043] As a further aspect of the present invention, the step of comparing the weighted risk vector with a preset construction progress benchmark matrix and generating a resource scheduling correction instruction specifically includes:

[0044] Read the pre-stored construction progress baseline matrix from the project management database and extract the baseline progress row vector corresponding to the current decision cycle from the matrix;

[0045] The weighted risk vector is subtracted element by element from the baseline progress row vector to obtain the progress deviation vector for each construction area.

[0046] Perform a sign determination operation on each deviation value in the schedule deviation vector and mark the construction area with the deviation value exceeding the preset risk threshold as a high-risk area;

[0047] For each construction area marked as a high-risk area, extract the current resource occupancy and the current task completion percentage from its corresponding area state vector;

[0048] Calculate the additional resources required for the high-risk area based on the current resource usage, the current task completion percentage, and the deviation value in the progress deviation vector;

[0049] All high-risk areas and the corresponding resource additions for each high-risk area are organized into a resource scheduling correction instruction list;

[0050] The list of resource scheduling correction instructions is sent to the resource scheduling execution unit of the hydropower project.

[0051] As a further aspect of the present invention, the step of performing a sign determination operation on each deviation value in the progress deviation vector specifically includes: if the deviation value is positive, it indicates that the progress of the corresponding construction area is lagging behind the baseline plan; if the deviation value is negative, it indicates that the progress of the corresponding construction area is ahead of the baseline plan.

[0052] As a further aspect of the present invention, the present invention also includes an artificial intelligence-based hydropower project management system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the artificial intelligence-based hydropower project management method described above.

[0053] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0054] By using graph neural networks to model the original temporal feature sets of each construction area in a hydropower project, the inherent spatial dependencies between different construction areas are mined, generating a structured regional association feature map. This changes the isolated data analysis approach for each area, incorporating implicit connections such as process associations, location associations, and resource associations between areas into the feature modeling scope. It fully preserves the topological association information between multiple construction areas, providing a stable spatial feature foundation for subsequent temporal feature mining.

[0055] Temporal convolution operations are performed on the regional correlation feature map to extract continuous change features of construction status within multiple time windows. Spatial correlation structure and temporal evolution information are integrated to construct a dimensionally regular dynamic state tensor. This overcomes the limitations of static analysis at a single moment, achieving a deep representation of the long-term change law of construction status and carrying the overall state information of project operation across multiple time periods.

[0056] By inputting dynamic state tensors into a multi-head attention mechanism, the system autonomously calculates the risk weight distribution of each construction area within the decision-making cycle, outputting a quantified weighted risk vector. This approach breaks away from uniform and rigid risk rating methods, automatically allocating risk proportions based on the actual operational status of each area, resulting in a differentiated quantitative risk expression. The weighted risk vector is then compared item by item with the construction progress baseline matrix, and resource scheduling correction instructions are directly generated based on the deviation results. This completes a closed-loop process from project risk quantification assessment to dynamic resource allocation, meeting the business needs of hydropower projects for routine intelligent analysis, quantitative risk management, and refined dynamic resource allocation. Attached Figure Description

[0057] Figure 1 This is a state diagram of the artificial intelligence-based hydropower project management method described in this invention;

[0058] Figure 2 A flowchart illustrating the workflow for collecting real-time data from hydropower project construction areas to construct the original time-series feature set;

[0059] Figure 3 A flowchart illustrating the workflow for extracting region association feature maps for neural networks. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] See Figure 1 This invention provides an artificial intelligence-based hydropower project management method, the specific method including:

[0062] Real-time monitoring data from each construction area of ​​the hydropower project is collected to form an initial temporal feature set. This initial temporal feature set is then input into a graph neural network to extract the spatial dependencies between construction areas and generate a regional correlation feature map. Temporal convolution operations are performed on this regional correlation feature map to capture the state change patterns of each construction area under different time windows and form a dynamic state tensor. This dynamic state tensor is then passed to a multi-head attention mechanism to calculate the risk weight distribution of each construction area within the current decision-making cycle and output a weighted risk vector. Based on this weighted risk vector, a difference comparison is performed with a preset construction progress benchmark matrix, and resource scheduling correction instructions are generated.

[0063] In one embodiment of the present invention, when collecting real-time monitoring data from various construction areas of a hydropower project and constructing an original time-series feature set, refer to... Figure 2 Multiple sensor nodes are deployed in each construction area of ​​the hydropower project to synchronously collect environmental parameter data and equipment operating status data. A unified timestamp is added to the environmental parameter data and equipment operating status data collected by each sensor node, and the data with the unified timestamp is transmitted to an edge computing gateway. At the edge computing gateway, the data from different sensor nodes are grouped and aggregated according to the construction area number. Missing value detection is performed on each grouped and aggregated data, and missing values ​​are filled using the mean of data at adjacent timestamps within the same construction area. This missing value detection operation specifically includes using the isolated forest algorithm to identify abnormal missing values ​​caused by sensor malfunctions. The data for each construction area after missing value filling is arranged into a time series in ascending order of timestamps. The time series of all construction areas are aligned by construction area number and construction area name and combined into a multidimensional array as the original time series feature set.

[0064] In the specific implementation, for five construction areas of a hydropower project—the dam pouring area, the water diversion tunnel area, the underground powerhouse area, the floodgate area, and the sand and gravel processing area—multiple sensor nodes are deployed in each construction area to synchronously collect environmental parameter data and equipment operating status data. The environmental parameter data includes temperature, humidity, and rainfall, while the equipment operating status data includes the vibration frequency and current load of the concrete mixer, the lifting weight and rotation angle of the tower crane, and the hydraulic pressure and fuel consumption rate of the excavator. A unified timestamp is added to the environmental parameter data and equipment operating status data collected by each sensor node, and the data with the unified timestamp is transmitted to the edge computing gateway. The timestamp uses the UTC time format with millisecond precision. In some embodiments, at the edge computing gateway, data from different sensor nodes are grouped and aggregated according to construction area numbers. The construction area numbers are: dam pouring area (Z01), water diversion tunnel area (Z02), underground powerhouse area (Z03), floodgate area (Z04), and sand and gravel processing area (Z05). Data collected by all sensor nodes within each construction area is aggregated into a data group named after that construction area number. In some embodiments, a missing value detection operation is performed on each grouped and aggregated data, and missing values ​​are filled using the average data from the same construction area at adjacent timestamps. Specifically, the missing value detection operation includes using an isolated forest algorithm to identify anomalous missing values ​​caused by sensor malfunctions. In the isolated forest algorithm, the subsampling size is set to 256 trees, the number of trees is 100, and the anomaly score threshold is set to 0.6. When the anomaly score of a data point exceeds 0.6, that data point is determined to be an anomalous missing value. For data points determined to be anomalous missing values, the average data from the same construction area at adjacent timestamps is used for filling, with the filling formula being:

[0065]

[0066] in: This represents the data after filling in the missing value for construction area number i at timestamp t. This represents data collected at timestamp t-1 from the same construction area. This represents data collected at timestamp t+1 from the same construction area. The data for each construction area, after filling in missing values, is arranged into a time series sequence in ascending order of timestamps. Taking the dam pouring area Z01 as an example, this time series sequence sequentially contains data vectors for nine parameter dimensions—temperature, humidity, rainfall, vibration frequency, current load, lifting weight, rotation angle, hydraulic pressure, and fuel consumption rate—at each hour from the 1st to the 720th hour of the construction start date. The time series sequences of all construction areas are aligned according to the construction area number and name and combined into a multi-dimensional array as the original time series feature set. The dimensions of this multi-dimensional array are the number of construction areas (5), the time step (720), and the parameter dimension (9). Optionally, after performing grouping and aggregation at the edge computing gateway, a duplicate data deletion operation is performed on the data groups of each construction area, deleting redundant records with identical timestamps and all identical parameter values. Optionally, when using the isolated forest algorithm to identify abnormal missing values, the data groups of each construction area are pre-standardized, scaling the values ​​of each parameter dimension to a range with a mean of 0 and a standard deviation of 1. It is understood that adjacent timestamps refer to two timestamps that are directly before and after the missing timestamp in chronological order. If the missing timestamp is at the beginning of the time sequence, it is filled using only the data of the next timestamp; if the missing timestamp is at the end of the time sequence, it is filled using only the data of the previous timestamp. It is also understood that when the multidimensional array is combined, the time sequence of each construction area is arranged sequentially in the first dimension of the array according to the construction area numbers Z01 to Z05. Within each construction area, the time sequence is arranged in the second dimension of the array according to the timestamps from smallest to largest. The data vector under each timestamp is arranged in the third dimension of the array according to the order of environmental parameter data first, followed by equipment operating status data.

[0067] In one embodiment of the present invention, when the original temporal feature set is input into a graph neural network to extract the spatial dependencies between construction areas and generate a regional association feature map, see [reference needed]. Figure 3Each construction area is treated as a graph node, and the temporal sequence of each construction area in the original temporal feature set is used as the initial feature vector of that graph node. The physical connections between graph nodes are determined based on the overall construction layout plan of the hydropower project, and an undirected edge is constructed between any two physically connected graph nodes. All graph nodes and all undirected edges are combined into an initial graph structure. This initial graph structure and the initial feature vector of each graph node are input into the message passing layer of a graph neural network, which is a graph attention network. In the message passing layer of the graph neural network, the feature vectors of each graph node's neighboring graph nodes are aggregated to update the feature representation of that graph node. Multiple message passing operations are performed to ensure that feature information is fully diffused among all graph nodes globally. After message passing is completed, the updated feature vectors corresponding to each graph node are arranged in the order of the graph nodes into a two-dimensional matrix as the associated feature map of that region.

[0068] In specific implementation, for a hydropower project with five construction areas—dam pouring area, water diversion tunnel area, underground powerhouse area, flood discharge gate area, and sand and gravel processing area—each construction area is treated as a graph node. The time sequence of each construction area in the original time sequence feature set generated in the above embodiment is used as the initial feature vector of the graph node. The time sequence of the dam pouring area is 720 hours long and contains nine parameter dimensions. The time sequences of the water diversion tunnel area, underground powerhouse area, flood discharge gate area, and sand and gravel processing area have the same dimensional structure. The physical connection relationships between each graph node are determined based on the overall construction layout plan of the hydropower project. An undirected edge is constructed between two graph nodes that have a physical connection. In the overall construction layout diagram, there are transportation roads connecting the dam pouring area and the water diversion tunnel area; a conveyor belt connecting the dam pouring area and the sand and gravel processing area; a construction access road connecting the water diversion tunnel area and the underground powerhouse area; and a temporary road connecting the underground powerhouse area and the floodgate area. Therefore, an undirected edge is constructed between the graph nodes in the dam pouring area and the water diversion tunnel area; between the graph nodes in the dam pouring area and the sand and gravel processing area; between the graph nodes in the water diversion tunnel area and the underground powerhouse area; and between the graph nodes in the underground powerhouse area and the floodgate area. All graph nodes and all undirected edges are combined into an initial graph structure, which contains five graph nodes and four undirected edges.

[0069] In some embodiments, the initial graph structure and the initial feature vector of each graph node are input into the message passing layer of the graph neural network, which is a graph attention network. In the message passing layer of the graph neural network, the feature vectors of the neighboring graph nodes of each graph node are aggregated to update the feature representation of the graph node. Specifically, the attention coefficient calculation mechanism in the graph attention network is used to weight and aggregate the features of the neighboring graph nodes. For any graph node j and its neighboring graph node k, the attention coefficient calculation formula is:

[0070]

[0071] in: This represents the unnormalized attention score between graph node j and its neighboring graph node k in the l-th layer of the graph neural network. This represents the transpose of the learnable attention weight vector in the l-th layer of the graph neural network. This represents the learnable linear transformation weight matrix in the l-th layer of the graph neural network. This represents the feature vector of input graph node j in the l-th layer of the graph neural network. This represents the feature vector of the input neighbor graph node k in the l-th layer of the graph neural network. The vector concatenation operation is represented by `LeakyReLU`, which represents a linear rectified activation function with leakage. The unnormalized attention score is normalized using a softmax function to obtain normalized attention weights. These normalized attention weights are then used to weight and sum the transformed features of neighboring graph nodes to update the feature representation of graph node `j`. The message passing operation is repeated three times to ensure sufficient diffusion of feature information among global graph nodes. In the first message passing operation, each graph node aggregates features only from its directly connected neighboring graph nodes; in the second, it aggregates features from nodes within a two-hop distance; and in the third, it aggregates features from nodes within a three-hop distance. Optionally, four attention heads are set in each layer of the graph attention network. Each attention head independently calculates attention coefficients and generates corresponding updated features. The updated features generated by the four attention heads are concatenated along the feature dimension to serve as the current layer output feature of the graph node. Optionally, after each message passing operation, a random dropout operation with a dropout rate of 0.2 is performed on the updated features of each graph node to prevent overfitting.

[0072] It is understandable that the construction of undirected edges is based solely on the physical connections marked on the overall construction layout plan. If two construction areas do not have direct physical connections such as roads, conveyor belts, or pipelines on the overall construction layout plan, then no undirected edge is constructed between the corresponding two graph nodes. It is also understandable that the number of repeated message passing operations is equal to the number of layers in the graph neural network. Setting the number of layers to three allows each graph node to obtain the feature information of all other graph nodes in the entire initial graph structure, since the diameter of the initial graph structure does not exceed three hops. After message passing is completed, the updated feature vectors corresponding to each graph node are arranged into a two-dimensional matrix according to the graph node order as a region association feature map. The graph node order is: dam pouring area, water diversion tunnel area, underground powerhouse area, floodgate area, and sand and gravel processing area. The two-dimensional matrix has five rows and the number of columns is the dimension length of each updated feature vector.

[0073] In one embodiment of the present invention, when performing temporal convolution operations on the region-related feature map to capture the state change patterns of each construction area under different time windows and form a dynamic state tensor, the region-related feature map is divided into multiple continuous and non-overlapping fixed-length time windows along the time dimension. An independent temporal convolution kernel is configured for each fixed-length time window, and the size of each temporal convolution kernel is set to be equal to the length of the time window. Each temporal convolution kernel is element-wise multiplied with the region-related feature map within its corresponding fixed-length time window, and then summed to obtain the convolution output feature of that time window. The convolution output features of all time windows are reassembled into a three-dimensional feature tensor in chronological order. Batch normalization is performed on this three-dimensional feature tensor to eliminate dimensional differences between different construction areas. The three-dimensional feature tensor after batch normalization is used as the dynamic state tensor.

[0074] In specific implementation, for a certain hydropower project, there are five construction areas: the dam pouring area, the water diversion tunnel area, the underground powerhouse area, the floodgate area, and the sand and gravel processing area. The regional association feature map generated in the above embodiment is divided into multiple continuous and non-overlapping fixed-length time windows along the time dimension. The size of the regional association feature map is 5 times the number of construction areas, 720 times the time step, and 64 times the output feature dimension of the graph neural network. The length of each fixed-length time window is set to 24 time steps. There is no overlap between adjacent time windows. Therefore, the 720 time steps are divided into 30 continuous and non-overlapping time windows. Each time window covers the construction period from hour 1 to 24, hour 25 to 48, ... up to hour 697 to 720. In some embodiments, an independent temporal convolutional kernel is configured for each fixed-length time window, and the size of each temporal convolutional kernel is set to be equal to the length of the time window. Each temporal convolutional kernel is a three-dimensional tensor with a size equal to the number of construction areas (5), the time window length (24), and the feature dimension (64). 30 time windows correspond to 30 independent temporal convolutional kernels, and the parameters of each temporal convolutional kernel are updated independently during training. Within each fixed-length time window, the corresponding temporal convolutional kernel is multiplied element-wise with the sub-tensor covered by that time window in the region association feature map, and then summed to obtain the convolutional output feature of the time window. For the w-th time window, the formula for calculating the convolutional output feature is:

[0075]

[0076] in: The convolutional output feature at the w-th time window is a scalar value. The value range is from 1 to 30. The index represents the construction area, with a value ranging from 1 to 5. This represents the relative time step index within the time window, with a value ranging from 1 to 24. This represents the index of the feature dimension, with values ​​ranging from 1 to 64. This represents the parameter values ​​of the temporal convolution kernel corresponding to the w-th time window at the construction region index i, relative time step t, and feature dimension d. This represents the value of the sub-tensor covered by the w-th time window in the region association feature map at the construction region index i, relative time step t, and feature dimension d. The convolutional output features of all time windows are reassembled into a three-dimensional feature tensor in chronological order. This three-dimensional feature tensor has a size of 30 x 1 x 1, where the first dimension corresponds to the order of the 30 time windows, and the last two dimensions are both 1. In some embodiments, batch normalization is performed on the three-dimensional feature tensor to eliminate dimensional differences between different construction regions. The batch normalization operation calculates the mean and variance of the convolutional output features within each batch along the time window dimension, and uses the mean and variance to normalize each element in the three-dimensional feature tensor. After normalization, it is multiplied by a learnable scaling parameter and a learnable translation parameter is added. The initial value of the scaling parameter is set to 1, and the initial value of the translation parameter is set to 0. The three-dimensional feature tensor after batch normalization is used as a dynamic state tensor, and the size of the dynamic state tensor remains 30 x 1 x 1. Optionally, before concatenating the convolutional output features into a 3D feature tensor, a linear rectified activation function is performed on the convolutional output features of each time window, setting negative values ​​to zero. Optionally, a random dropout operation is performed after batch normalization, with a dropout rate set to 0.1. It can be understood that each time step in a continuous and non-overlapping fixed-length time window corresponds to a one-hour construction data sampling interval, and a time window length of 24 indicates that the state change pattern is extracted with a 24-hour cycle. It can be understood that the size of the temporal convolutional kernel being equal to the time window length means that each temporal convolutional kernel does not perform a sliding stride operation, and each time window only produces one convolutional output feature value.

[0077] In one embodiment of the present invention, when the dynamic state tensor is passed to a multi-head attention mechanism to calculate the risk weight distribution of each construction area in the current decision-making period and output a weighted risk vector, the dynamic state tensor is split into multiple independent region state vectors along the construction area dimension. In each attention head, a linear transformation is performed on each region state vector to generate the query vector, key vector, and value vector for that region. In each attention head, the similarity score between the query vector of each region and the key vectors of all regions is calculated, and a normalized exponential function is used to convert the similarity score into attention weights. In each attention head, the value vectors of all regions are weighted and summed according to their corresponding attention weights to generate the risk weighted vector for that attention head. The risk weighted vectors generated by all attention heads are concatenated along the feature dimension and then subjected to a linear transformation to generate a fused risk weighted vector. Each element in the fused risk weighted vector is used as the risk weight of the corresponding construction area in the current decision-making period, and the risk weights of all construction areas are arranged in the region index order to form the weighted risk vector.

[0078] In specific implementation, for a hydropower project encompassing five construction areas—dam pouring area, water diversion tunnel area, underground powerhouse area, flood discharge gate area, and sand and gravel processing area—the dynamic state tensor generated in the above embodiment is decomposed along the construction area dimension into multiple independent region state vectors. The size of the dynamic state tensor is 30 times the number of time windows, 5 times the number of construction areas, and 1 times the feature dimension. After decomposition along the construction area dimension, five independent region state vectors are obtained, each with a size of 30 times 1, corresponding to the state feature sequences of the dam pouring area, water diversion tunnel area, underground powerhouse area, flood discharge gate area, and sand and gravel processing area over the past 30 time windows (each time window is 24 hours, for a total of 720 hours). In some embodiments, a linear transformation is performed on each region state vector in each attention head to generate a query vector, key vector, and value vector for the region. The number of attention heads is set to 8, and an independent linear transformation matrix is ​​configured for each attention head. , , ,in The index of the attention head ranges from 1 to 8. The linear transformation matrix has a dimension of 1 x 64. Each region's state vector, multiplied by the transformation matrix, yields a query vector, key vector, and value vector with dimensions 30 x 64. Within each attention head, a similarity score is calculated between the query vector of each region and the key vectors of all regions. A normalized exponential function is used to convert the similarity score into attention weights. For the h-th attention head, any two construction regions... and The similarity score between them is calculated by dividing the construction area Query vector and construction area The key vector is divided by the scaling factor after performing a dot product operation on the feature dimension. After obtaining the similarity score matrix, for each construction area Perform a normalized exponential function transformation along the construction area dimension, so that all construction areas are relative to the construction area. The sum of attention weights is 1.

[0079] In each attention head, the value vectors of all regions are summed according to their corresponding attention weights to generate a risk-weighted vector for that attention head. For the h-th attention head, the construction region... The risk-weighted vector is the sum of the value vectors of all construction areas multiplied by their corresponding attention weights. Its calculation formula is as follows:

[0080]

[0081] in: This indicates that the h-th attention head output is for the construction area. The risk-weighted vector, The index represents the construction area, with a value ranging from 1 to 5. This indicates the construction area in the h-th attention head. For the construction area Normalized attention weights This indicates the construction area in the h-th attention head. The value vector has a dimension of 30 x 64. The risk weighted vectors generated by all attention heads are concatenated along the feature dimensions and then subjected to a linear transformation to generate a fused risk weighted vector. Each of the risk weighted vectors output by the eight attention heads has a dimension of 30 x 64. After concatenation, a tensor with a dimension of 30 x 512 is obtained. The linear transformation uses a learnable weight matrix with a dimension of 512 x 1 to map the concatenation result to a tensor of 30 x 1, which serves as the fused risk weighted vector. Each element in the fused risk weighted vector is used as the risk weight of the corresponding construction area within the current decision-making period. The risk weights of all construction areas are arranged in order of their region indices to form a weighted risk vector. The size of the fused risk weighted vector is 30 multiplied by 1, where 30 corresponds to the time window index and 1 corresponds to the construction area index. Since the current decision-making period corresponds to the last time window (the 30th time window), the value in the 30th row and first column of the fused risk weighted vector is taken as the risk weight for the dam pouring area. Similarly, five risk weight values ​​are extracted according to the order of the construction areas: dam pouring area, water diversion tunnel area, underground powerhouse area, floodgate area, and sand and gravel processing area. These five risk weight values ​​are arranged into a one-dimensional vector of length 5 as the weighted risk vector. Optionally, before performing a linear transformation to generate the fused risk weighted vector, a layer normalization operation is performed on the risk weighted vectors output by the eight attention heads. The layer normalization calculates the mean and variance along the feature dimension. Optionally, before calculating the similarity score in each attention head, a random discard operation with a discard rate of 0.1 is performed on the query vector and key vector. It is understood that the current decision-making cycle refers to the construction period covered by the most recently completed time window, i.e., the 697th to the 720th hour. It is also understood that the order of the construction area index is consistent with the order of the nodes in the above embodiment, which is dam pouring area, water diversion tunnel area, underground powerhouse area, floodgate area, and sand and gravel processing area, to ensure the correct correspondence between the weighted risk vector and the areas in subsequent resource scheduling correction instructions.

[0082] In one embodiment of the present invention, when generating resource scheduling correction instructions by comparing the weighted risk vector with a preset construction progress benchmark matrix, the pre-stored construction progress benchmark matrix is ​​read from the project management database, and the benchmark progress row vector corresponding to the current decision cycle is extracted from the matrix. The weighted risk vector is subtracted element-wise from the benchmark progress row vector to obtain the progress deviation vector for each construction area. A sign judgment operation is performed on each deviation value in the progress deviation vector, and construction areas with deviation values ​​exceeding a preset risk threshold are marked as high-risk areas. A positive deviation value indicates that the progress of the corresponding construction area is lagging behind the benchmark plan, and a negative deviation value indicates that the progress of the corresponding construction area is ahead of the benchmark plan. For each construction area marked as a high-risk area, the current resource occupancy and current task completion percentage are extracted from its corresponding area status vector. The required additional resources for the high-risk area are calculated based on the current resource occupancy, the current task completion percentage, and the deviation value in the progress deviation vector. All high-risk areas and the corresponding additional resources for each high-risk area are organized into a resource scheduling correction instruction list. This resource scheduling correction instruction list is sent to the resource scheduling execution unit of the hydropower project.

[0083] In practical implementation, for a hydropower project encompassing five construction areas—dam construction area, water diversion tunnel area, underground powerhouse area, floodgate area, and sand and gravel processing area—a pre-stored construction progress baseline matrix is ​​retrieved from the project management database. The baseline progress row vector corresponding to the current decision-making cycle is extracted from this matrix. The row index of the construction progress baseline matrix corresponds to the decision-making cycle number, and the column index corresponds to the construction area number. The current decision-making cycle is the 30th time window (corresponding to the 697th to 720th hour). The 30th row vector in the construction progress baseline matrix is... Where 0.45 represents the planned completion percentage of the dam pouring area at the end of 720 hours as 45%, 0.52 represents the planned completion percentage of the water diversion tunnel area as 52%, 0.38 represents the planned completion percentage of the underground powerhouse area as 38%, 0.61 represents the planned completion percentage of the flood discharge gate area as 61%, and 0.73 represents the planned completion percentage of the sand and gravel processing area as 73%. In some embodiments, the weighted risk vector is subtracted element-wise from the baseline progress row vector to obtain the progress deviation vector for each construction area. The weighted risk vector is generated by the above embodiment, and its value is... Subtracting each element one by one yields the progress deviation vector. Positive deviation values ​​indicate that the actual progress lags behind the baseline plan, while negative deviation values ​​indicate that the actual progress is ahead of the baseline plan. A sign judgment operation is performed on each deviation value in the progress deviation vector, and construction areas with deviation values ​​exceeding a preset risk threshold are marked as high-risk areas. The preset risk threshold is set to 0.03. For the dam pouring area, the deviation value is 0.06, which is greater than 0.03. For the water diversion tunnel area, the absolute value of the deviation value -0.04 is 0.04, which is greater than 0.03, but this negative deviation value indicates that the progress is ahead of schedule and is therefore not marked as a high-risk area. For the underground powerhouse area, the deviation value is 0.04, which is greater than 0.03. For the floodgate area, the deviation value is 0.04, which is greater than 0.03. For the sand and gravel processing area, the absolute value of the deviation value -0.03 is equal to 0.03, which does not exceed the threshold and is negative, so it is not marked. Therefore, the construction areas marked as high-risk areas include the dam pouring area, the underground powerhouse area, and the floodgate area.

[0084] For each construction area marked as a high-risk area, the current resource occupancy and current task completion percentage are extracted from its corresponding area state vector. The area state vector is derived from the independent area state vector obtained by splitting along the construction area dimension in the above embodiment. Specifically, the current resource occupancy of the dam pouring area is 3 concrete mixers, 2 tower cranes, and 45 construction workers, with a current task completion percentage of 39%; the current resource occupancy of the underground powerhouse area is 2 excavators, 1 loader, and 28 construction workers, with a current task completion percentage of 34%; and the current resource occupancy of the floodgate area is 4 vibrators and 22 construction workers, with a current task completion percentage of 57%. The required additional resources for the high-risk area are calculated based on the current resource occupancy, current task completion percentage, and deviation value in the progress deviation vector. The formula for calculating the additional resources is as follows:

[0085]

[0086] in: This represents the additional resources required for construction area j (measured in standard resource units, one standard resource unit being equivalent to the labor of one construction worker working for one hour). The values ​​are indexes for high-risk areas, corresponding to the dam construction area, underground powerhouse area, and floodgate area, respectively. This represents the progress deviation value (absolute value, i.e., lag) for construction area j. This represents the current resource usage in construction area j (after conversion to standard resource units). This represents the current percentage of task completion (expressed as a decimal) for construction area j. For the dam pouring area, the schedule deviation value is... The current resource usage is 0.06. The project involves 3 concrete mixers (each equivalent to 15 standard resource units), 2 tower cranes (each equivalent to 20 standard resource units), and 45 construction workers (each equivalent to 1 standard resource unit), totaling 3×15+2×20+45=45+40+45=130 standard resource units. The current task completion percentage is... The calculated resource addition amount is 0.39. Standard resource unit. For the underground powerhouse area, schedule deviation value. The current resource usage is 0.04. The project involves 2 excavators (each equivalent to 25 standard resource units), 1 loader (each equivalent to 18 standard resource units), and 28 construction workers (each equivalent to 1 standard resource unit), totaling 2×25+1×18+28=50+18+28=96 standard resource units. The current task completion percentage is... The calculated resource addition amount is 0.34. Standard resource unit. For flood discharge gate areas, schedule deviation value. The current resource usage is 0.04. The project involves 4 vibratory compactors (each equivalent to 8 standard resource units) and 22 construction workers (each equivalent to 1 standard resource unit), totaling 4 × 8 + 22 = 32 + 22 = 54 standard resource units. The current task completion percentage is as follows: The calculated resource addition amount is 0.57. Standard resource units. Optionally, when the schedule deviation is negative (indicating schedule overrun), no additional resources are calculated; instead, a resource release instruction is generated. The formula for calculating the resource release amount is the same as the formula above, but with a negative sign. All high-risk areas and their corresponding additional resource amounts are organized into a resource scheduling correction instruction list. This list contains three records: the first record is an additional resource amount of 12.79 standard resource units for the dam pouring area; the second record is an additional resource amount of 5.82 standard resource units for the underground powerhouse area; and the third record is an additional resource amount of 5.02 standard resource units for the spillway area. The resource scheduling correction instruction list is sent to the hydropower project's resource scheduling execution unit. Based on the instruction list, the resource scheduling execution unit allocates the corresponding number of standard resource units from the overrunning water diversion tunnel area and sand and gravel processing area to the three high-risk areas. It can be understood that positive deviation values ​​in the schedule deviation vector indicate that the corresponding construction area is lagging behind the baseline plan, while negative deviation values ​​indicate that the corresponding construction area is ahead of the baseline plan. The sign judgment operation only marks construction areas with positive deviation values ​​that exceed the preset risk threshold as high-risk. The conversion factor of equipment to standard resource units in the current resource occupancy is pre-stored in the project management database. The conversion factor is determined based on the rated power of the equipment and the energy consumption per unit time.

[0087] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A hydropower project management method based on artificial intelligence, characterized in that: include: Collect real-time monitoring data from various construction areas of the hydropower project and construct an original time-series feature set; The original temporal feature set is input into a graph neural network to extract the spatial dependencies between construction areas and generate a regional association feature map. Temporal convolution operations are performed on the region association feature map to capture the state change patterns of each construction area under different time windows and form a dynamic state tensor. The dynamic state tensor is passed to the multi-head attention mechanism to calculate the risk weight distribution of each construction area in the current decision-making cycle and output the weighted risk vector. The weighted risk vector is compared with the preset construction progress benchmark matrix to generate a resource scheduling correction instruction.

2. The artificial intelligence-based hydropower project management method according to claim 1, characterized in that, The steps for collecting real-time monitoring data from various construction areas of a hydropower project and constructing an original time-series feature set specifically include: Multiple sensor nodes are deployed in each construction area of ​​the hydropower project to synchronously collect environmental parameter data and equipment operating status data; Add a unified timestamp to the environmental parameter data and equipment operating status data collected by each sensor node, and then transmit the data with the unified timestamp to the edge computing gateway; At the edge computing gateway, data from different sensor nodes are grouped and aggregated according to the construction area number; For each grouped aggregated data, a missing value detection operation is performed, and missing values ​​are filled using the mean of data from the same construction area at adjacent timestamps. After filling in the missing values, the data for each construction area are arranged into a time series in ascending order of timestamps. The time series sequences of all construction areas are aligned according to the construction area number and construction area name and combined into a multidimensional array as the original time series feature set.

3. The artificial intelligence-based hydropower project management method according to claim 2, characterized in that, The missing value detection operation specifically includes using the isolated forest algorithm to identify abnormal missing values ​​caused by sensor malfunctions.

4. The artificial intelligence-based hydropower project management method according to claim 2, characterized in that, The step of inputting the original temporal feature set into a graph neural network to extract the spatial dependencies between construction areas and generate regional association feature maps specifically includes: Each construction area is taken as a graph node, and the time sequence of each construction area in the original time sequence feature set is taken as the initial feature vector of the graph node; Based on the overall construction layout plan of the hydropower project, determine the physical connection relationship between each node in the plan and construct an undirected edge between two nodes that have a physical connection relationship. Combine all graph nodes and all undirected edges into an initial graph structure; The initial graph structure and the initial feature vector of each graph node are input into the message passing layer of the graph neural network; In the message-passing layer of the graph neural network, the feature vectors of the neighboring graph nodes of each graph node are aggregated to update the feature representation of that graph node; Repeatedly executing message passing operations allows feature information to be fully diffused among nodes in the global graph. After message passing is completed, the updated feature vectors corresponding to each graph node are arranged into a two-dimensional matrix according to the graph node order, which serves as the region association feature map.

5. The artificial intelligence-based hydropower project management method according to claim 4, characterized in that, The graph neural network is a graph attention network.

6. The artificial intelligence-based hydropower project management method according to claim 4, characterized in that, The steps of performing temporal convolution operations on the region association feature map to capture the state change patterns of each construction area under different time windows and form a dynamic state tensor specifically include: The region association feature map is divided into multiple continuous and non-overlapping fixed-length time windows along the time dimension; Configure an independent temporal convolution kernel for each fixed-length time window and set the size of each temporal convolution kernel to be equal to the length of the time window; The convolutional output features of that time window are obtained by multiplying each temporal convolution kernel element-wise with the region-related feature map within its corresponding fixed-length time window and then summing the results. The convolutional output features of all time windows are reassembled into a three-dimensional feature tensor in chronological order. Batch normalization is performed on the three-dimensional feature tensor to eliminate dimensional differences between different construction areas; The three-dimensional feature tensor after batch normalization is used as the dynamic state tensor.

7. The artificial intelligence-based hydropower project management method according to claim 6, characterized in that, The steps of passing the dynamic state tensor to the multi-head attention mechanism to calculate the risk weight distribution of each construction area within the current decision-making cycle and outputting the weighted risk vector specifically include: The dynamic state tensor is split into multiple independent region state vectors along the construction region dimension. In each attention head, a linear transformation is performed on the state vector of each region to generate the query vector, key vector, and value vector for that region. In each attention head, the similarity score between the query vector of each region and the key vector of all regions is calculated, and the similarity score is converted into attention weights using a normalized exponential function; In each attention head, the value vectors of all regions are summed according to their corresponding attention weights to generate the risk-weighted vector for that attention head; After concatenating the risk weighted vectors generated by all attention heads along the feature dimension, a linear transformation is performed to generate a fused risk weighted vector; Each element in the fused risk weighted vector is used as the risk weight of the corresponding construction area in the current decision-making cycle, and the risk weights of all construction areas are arranged in the order of the area index to form the weighted risk vector.

8. The artificial intelligence-based hydropower project management method according to claim 7, characterized in that, The step of comparing the weighted risk vector with the preset construction progress benchmark matrix and generating resource scheduling correction instructions specifically includes: Read the pre-stored construction progress baseline matrix from the project management database and extract the baseline progress row vector corresponding to the current decision cycle from the matrix; The weighted risk vector is subtracted element by element from the baseline progress row vector to obtain the progress deviation vector for each construction area. Perform a sign determination operation on each deviation value in the schedule deviation vector and mark the construction area with the deviation value exceeding the preset risk threshold as a high-risk area; For each construction area marked as a high-risk area, extract the current resource occupancy and the current task completion percentage from its corresponding area state vector; Calculate the additional resources required for the high-risk area based on the current resource usage, the current task completion percentage, and the deviation value in the progress deviation vector; All high-risk areas and the corresponding resource additions for each high-risk area are organized into a resource scheduling correction instruction list; The list of resource scheduling correction instructions is sent to the resource scheduling execution unit of the hydropower project.

9. The artificial intelligence-based hydropower project management method according to claim 8, characterized in that, The step of performing a sign determination operation on each deviation value in the schedule deviation vector specifically includes: if the deviation value is positive, it means that the progress of the corresponding construction area is lagging behind the baseline plan; if the deviation value is negative, it means that the progress of the corresponding construction area is ahead of the baseline plan.

10. An artificial intelligence-based hydropower project management system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based hydropower project management method according to any one of claims 1 to 9.

Images