Water conservancy and hydropower construction resource management scheduling method and system based on artificial intelligence

Abstract

The invention discloses a water conservancy and hydropower construction resource management scheduling method and system based on artificial intelligence. The method comprises the following steps: collecting construction multi-source data in real time; based on an ST-GNN space-time diagram neural network, predicting the resource demand of each construction area in a future T time period, and combining with an LSTM long short-term memory network to perform joint modeling on the construction progress and environmental disturbance to obtain a demand prediction model; adjusting an inertia weight and a cognitive factor of the demand prediction model by using an IPSO improved particle swarm optimization algorithm, and learning an optimal scheduling strategy in a simulation environment in combination with an RL reinforcement learning training agent to obtain a target demand prediction model; and inputting the construction multi-source data into the target demand prediction model to output a scheduling scheme, and performing digital twin simulation verification and adjustment on the scheduling scheme to obtain a target scheduling scheme. And the prediction accuracy of future resource demands and the management scheduling efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a water conservancy and hydropower construction resource management and scheduling method and system based on artificial intelligence. BACKGROUND

[0002] Water conservancy and hydropower projects usually have complex construction environments, large engineering scales, numerous participants, diverse resources, and frequent scheduling. Traditional resource management and scheduling methods rely heavily on the personal experience of project managers, static progress plans, and periodic reports, making it difficult to accurately predict future short-term resource needs for each work area, resulting in resource deployment lag or redundancy. The scheduling of various resources is often independent, lacking system-level collaborative optimization, and is prone to resource conflicts or resource idling. When encountering weather changes, geological condition mutations, equipment failures, and other disturbances, traditional methods are difficult to generate effective response plans in a timely manner, resulting in low scheduling efficiency. SUMMARY

[0003] The present application aims to solve the above problems and designs a water conservancy and hydropower construction resource management and scheduling method and system based on artificial intelligence.

[0004] To achieve the above purpose, the technical solution of the present application is as follows: further, in the above-mentioned water conservancy and hydropower construction resource management and scheduling method based on artificial intelligence, the water conservancy and hydropower construction resource management and scheduling method comprises the following steps:

[0005] Obtain Internet of Things data sensor data, BIM model data, GIS data, and construction record data during the water conservancy and hydropower construction process, preprocess the collected data, and obtain construction multi-source data;

[0006] Based on the ST-GNN spatio-temporal graph neural network, predict the resource demand of each construction area in the future T time period, combine the LSTM long short-term memory network to jointly model the construction progress and environmental disturbance, and obtain a demand prediction model;

[0007] Adjust the inertia weight and cognitive factor of the demand prediction model using the IPSO improved particle swarm optimization algorithm, combine the RL reinforcement learning to train the agent to learn the optimal scheduling strategy in the simulated environment, and obtain a target demand prediction model;

[0008] Input the construction multi-source data into the target demand prediction model to output a scheduling scheme, perform digital twin simulation verification and adjustment on the scheduling scheme, and obtain a target scheduling scheme.

[0009] Further, in the above-mentioned water conservancy and hydropower construction resource management and scheduling method based on artificial intelligence, the Internet of Things data sensor data, BIM model data, GIS data and construction record data in the water conservancy and hydropower construction process are obtained, the collected data is preprocessed, and construction multi-source data is obtained, including:

[0010] Obtain the Internet of Things data sensor data, BIM model data, GIS data and construction record data in the water conservancy and hydropower construction process;

[0011] Convert the sensor binary data into JSON format, convert the IFC format data of the BIM model into CSV structured data, convert the shp format data of the GIS into GeoJSON format, and unify the timestamp format;

[0012] Different methods are used for missing value processing according to different data types, sensor abnormal data is removed through the 3σ principle, and abnormal records are screened through business rules to obtain construction cleaning data;

[0013] PCA principal component analysis is used to reduce the dimension of high-dimensional data in the construction cleaning data to obtain construction multi-source data.

[0014] Further, in the above-mentioned water conservancy and hydropower construction resource management and scheduling method based on artificial intelligence, the resource demand of each construction area in the future T period is predicted based on the ST-GNN spatio-temporal graph neural network, and the construction progress and environmental disturbance are jointly modeled by combining the LSTM long short-term memory network to obtain a demand prediction model, including:

[0015] Divide the construction area into construction surfaces, resource supply points and transportation, and define them as graph nodes, construct edges according to the association relationship between nodes, and embed the time series data and spatial data of each node as node features into the spatio-temporal graph;

[0016] The ST-GNN spatio-temporal graph neural network includes a spatial convolution and a time convolution double module, the spatial convolution layer captures the resource flow association between nodes through the GAT graph attention mechanism, and the time convolution layer captures the time series trend through the 1D convolution.

[0017] Further, in the above-mentioned water conservancy and hydropower construction resource management and scheduling method based on artificial intelligence, the resource demand of each construction area in the future T period is predicted based on the ST-GNN spatio-temporal graph neural network, and the construction progress and environmental disturbance are jointly modeled by combining the LSTM long short-term memory network to obtain a demand prediction model, including:

[0018] A double-layer LSTM network is constructed, the first layer learns the time series rule of the progress feature, the second layer fuses the environmental disturbance feature, and outputs the construction progress prediction value in the future period and the influence coefficient of the environmental disturbance on the progress;

[0019] The node future period resource demand features output by the ST-GNN are spliced with the progress and environment joint features output by the LSTM,

[0020] The features are assigned weights by the attention layer, the LSTM feature weight is high for progress-sensitive resources, and the ST-GNN spatial feature weight is high for environment-sensitive resources.

[0021] Further, in the above-mentioned water conservancy and hydropower construction resource management and scheduling method based on artificial intelligence, the inertia weight and cognitive factor of the demand prediction model are adjusted by using the IPSO improved particle swarm optimization algorithm, and the agent is trained in the simulated environment to learn the optimal scheduling strategy by combining RL reinforcement learning, to obtain a target demand prediction model, comprising:

[0022] Initialize the particle swarm, calculate the fitness of each particle, adjust the particle position according to the inertia weight, cognitive factor and social factor, generate a new parameter combination, and output the optimal parameter combination when the global optimal fitness changes by less than or equal to 1% for 10 consecutive rounds.

[0023] Further, in the above-mentioned water conservancy and hydropower construction resource management and scheduling method based on artificial intelligence, the inertia weight and cognitive factor of the demand prediction model are adjusted by using the IPSO improved particle swarm optimization algorithm, and the agent is trained in the simulated environment to learn the optimal scheduling strategy by combining RL reinforcement learning, to obtain a target demand prediction model, comprising:

[0024] The state space is defined as a multi-dimensional vector of the current resource library, the demand prediction value of each region, the construction progress deviation, and the environmental disturbance coefficient, and the action space is defined as the resource scheduling action;

[0025] The DDPG deep deterministic policy gradient algorithm is used to construct the agent, the Actor network inputs the state and outputs a continuous scheduling action, and the Critic network evaluates the value of the action to obtain a target demand prediction model.

[0026] Further, in the above-mentioned water conservancy and hydropower construction resource management and scheduling method based on artificial intelligence, the construction multi-source data is input into the target demand prediction model to output a scheduling scheme, the scheduling scheme is verified and adjusted by digital twin simulation to obtain a target scheduling scheme, comprising:

[0027] The simulation results are compared with the preset target, the problem is traced back to the root cause by the twin model, an adjustment scheme is developed for the problem, and a target scheduling scheme is obtained.

[0028] Further, in the water conservancy and hydropower construction resource management and scheduling system based on artificial intelligence, the water conservancy and hydropower construction resource management and scheduling system comprises the following modules:

[0029] The construction data acquisition module is configured to acquire Internet of Things data sensor data, BIM model data, GIS data and construction record data in the water conservancy and hydropower construction process, pre-process the collected data, and obtain construction multi-source data.

[0030] The prediction model establishment module is configured to predict resource requirements of each construction area in a future T period based on an ST-GNN spatio-temporal graph neural network, combine an LSTM long short-term memory network to jointly model construction progress and environmental disturbance, and obtain a demand prediction model.

[0031] The prediction model optimization module is configured to adjust inertia weight and cognitive factor of the demand prediction model by using an IPSO improved particle swarm optimization algorithm, combine RL reinforcement learning to train an agent to learn an optimal scheduling strategy in a simulated environment, and obtain a target demand prediction model.

[0032] The scheduling scheme generation module is configured to input the construction multi-source data into the target demand prediction model to output a scheduling scheme, perform digital twin simulation verification and adjustment on the scheduling scheme, and obtain a target scheduling scheme.

[0033] Further, in the water conservancy and hydropower construction resource management and scheduling system based on artificial intelligence, the prediction model optimization module comprises the following sub-modules:

[0034] The definition sub-module is configured to define a state space as a multi-dimensional vector of a current resource library, a regional demand prediction value, a construction progress deviation and an environmental disturbance coefficient, and define an action space as a resource scheduling action.

[0035] The construction sub-module is configured to construct an agent by using a DDPG deep deterministic policy gradient algorithm, input a state into an Actor network to output a continuous scheduling action, and evaluate the value of the action by using a Critic network, so as to obtain the target demand prediction model.

[0036] Further, in the water conservancy and hydropower construction resource management and scheduling system based on artificial intelligence, the scheduling scheme generation module comprises the following sub-modules:

[0037] The obtaining sub-module is configured to compare simulation results with a preset target, backtrack a problem source by using a twin model, formulate an adjustment scheme for the problem, and obtain a target scheduling scheme.

[0038] Its beneficial effect lies in that by fusing ST-GNN and LSTM model, the spatio-temporal dynamic characteristics of resource demand and the long-term dependence relationship of progress, environment and other time sequence disturbance factors can be captured at the same time, the accuracy of future resource demand prediction and management and scheduling efficiency are improved, advanced and reliable data support is provided for scheduling decision, and passive response is changed into active intervention. The IPSO algorithm is used to adaptively adjust the parameters, which greatly improves the convergence speed and accuracy of the demand prediction model; and the RL reinforcement learning is used to explore the optimal strategy in the simulation environment, effectively solving the complex resource scheduling optimization problem under multi-objective and multi-constraint, and a global optimal or approximate optimal solution with the lowest cost, the highest efficiency and the most balanced resources can be generated. The environmental disturbance factors are embedded in the modeling process, and the scheduling scheme is simulated and verified and dynamically adjusted through digital twinning technology, so that the system can quickly respond and generate a robust scheduling scheme when facing uncertain interference such as weather and geological mutations, and the continuity and stability of construction are ensured. BRIEF DESCRIPTION OF DRAWINGS

[0039] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments with reference made to the accompanying drawings. The drawings are for purposes of illustration only and are not intended to be limiting of the present application.

[0040] Figure 1 The first embodiment schematic diagram of the water conservancy and hydropower construction resource management and scheduling method based on artificial intelligence in the embodiments of the present application;

[0041] Figure 2 The second embodiment schematic diagram of the water conservancy and hydropower construction resource management and scheduling method based on artificial intelligence in the embodiments of the present application;

[0042] Figure 3 The first embodiment schematic diagram of the water conservancy and hydropower construction resource management and scheduling system based on artificial intelligence in the embodiments of the present application. DETAILED DESCRIPTION

[0043] In order to make the objects, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0044] Those skilled in the art can understand that, unless specifically stated, the singular form "one", "an" and "the" used herein can also include the plural form. It should be further understood that the wording used in the specification of the present application includes the presence of a feature, integer, step, operation, element and / or component, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or their groups.

[0045] The application will be described in detail below with reference to the drawings, Figure 1 As shown, the artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method includes the following steps,

[0046] Step 101, obtain Internet of Things data sensor data, BIM model data, GIS data and construction record data in the water conservancy and hydropower construction process, preprocess the collected data to obtain construction multi-source data;

[0047] Specifically, in this embodiment, the Internet of Things data sensor data, BIM model data, GIS data and construction record data in the water conservancy and hydropower construction process are obtained; the sensor binary data is converted into JSON format, the IFC format data of the BIM model is converted into CSV structured data, the shp format data of the GIS is converted into GeoJSON format, and the timestamp format is unified; different data types are processed by a differentiated method, abnormal sensor data is removed by the 3σ principle, and abnormal records are screened by business rules to obtain construction cleaning data; PCA principal component analysis is used to reduce the dimension of high-dimensional data in the construction cleaning data to obtain construction multi-source data.

[0048] Internet of Things sensor data collection: special sensors are deployed for water conservancy and hydropower construction scenes, and the collected content includes: construction environment data; equipment operation data; resource state data. BIM model data extraction: structured data is exported from the water conservancy and hydropower construction BIM model, including: component parameters; progress correlation data; resource correlation data. GIS geographic spatial data collection: obtain geographic information of the construction area through the GIS system, including: terrain data; watershed data; traffic network data. Construction record data summary: collect construction data recorded by manual reporting and system automatic recording, including: personnel records; material consumption records; progress deviation records.

[0049] Format standardization for the heterogeneity of different source data: convert sensor binary data to JSON format, convert IFC format data of BIM model to CSV structured data, convert shp format data of GIS to GeoJSON format; unify timestamp format, such as YYYY-MM-DDHH:MM:SS, spatial coordinate system, use national 2000 geodetic coordinate system, resource measurement unit. Missing value processing, different methods are used for different data types: linear interpolation method is used for time series data such as water level and equipment oil consumption; for discrete data such as personnel attendance in construction records, the adjacent time period mean value is filled; for key data such as missing BIM component parameters, the construction drawings or field measurement are used to complete. Abnormal value processing: identify abnormalities according to the characteristics of water conservancy and hydropower construction scene: remove abnormal sensor data by 3σ principle; filter abnormal records by business rules. Redundant data elimination: delete duplicate data and data irrelevant to resource scheduling.

[0050] Construction of space-time-resource association database: time dimension: divide all data into day-week-month three-level time granularity according to construction progress plan, establish time sequence association; spatial dimension: map BIM component position, sensor deployment point, construction area, material yard / equipment storage point through GIS coordinates, realize spatial association of geographic location-resource-construction task; resource dimension: bind material, equipment, personnel data with construction tasks such as dam pouring and tunnel excavation, form task-resource-space three-dimensional data association table. Use principal component analysis, PCA to reduce the dimension of high-dimensional data: extract 3-5 principal components from 10+ dimensional equipment operation data such as speed, oil consumption and temperature, retain more than 90% information; for environmental data, remove redundant features through correlation analysis, such as when the correlation between wind speed and rainfall is >0.8, retain rainfall feature; finally output standardized construction multi-source data set, format is structured matrix of space-time feature+resource feature+progress feature+environment feature.

[0051] Step 102, based on ST-GNN space-time graph neural network, predict the resource demand of each construction area in the future T period, combine LSTM long short-term memory network to jointly model the construction progress and environmental disturbance, and obtain the demand prediction model;

[0052] Specifically, in the embodiment, the construction area is divided into a construction surface, a resource supply point and transportation, and is defined as a graph node. Time sequence data and spatial data of each node are embedded in the space-time graph as node features. The ST-GNN space-time graph neural network includes a space convolution and a time convolution double module. The space convolution layer captures the resource flow correlation between nodes through the GAT graph attention mechanism, and the time convolution layer captures the time sequence trend through the 1D convolution. A double-layer LSTM network is constructed. The first layer learns the time sequence rule of the progress feature, and the second layer fuses the environmental disturbance feature to output the construction progress prediction value in the future period and the influence coefficient of the environmental disturbance on the progress. The resource demand features of each node in the future period output by the ST-GNN are spliced with the progress and environmental joint features output by the LSTM, and the feature weights are distributed through the attention layer. The progress-sensitive resource is given a high LSTM feature weight, and the environment-sensitive resource is given a high ST-GNN space feature weight.

[0053] Construction area space-time graph construction: define graph nodes: divide the construction area into three types of nodes, including a construction surface such as a dam pouring area, a tunnel operation area, a resource supply point, a material yard, an equipment warehouse, a transportation node, a wharf and a convenient access point, a total of 30-50 nodes; define graph edges: construct edges according to the correlation between nodes, such as a transportation edge between a material yard and a construction surface, a weight of a transportation distance / time, and a process dependence edge between a construction surface and a construction surface, a weight of a progress correlation degree; space-time feature embedding: time sequence data of each node, such as resource demand in the past 7 days, and spatial data, such as node coordinates and resource storage capacity, are embedded in the space-time graph as node features. Space-time feature extraction and training: adopt the space convolution+time convolution double module of the ST-GNN: the space convolution layer captures the resource flow correlation between nodes through the graph attention mechanism, GAT, such as the supply influence of A material yard on B construction surface; the time convolution layer captures the time sequence trend through the 1D convolution, such as a weekly concrete demand peak; training target: minimize the mean square error, MSE, of the predicted node resource demand and the actual resource consumption of the model, and the number of iterations is set to 200 rounds and the batch size is 32.

[0054] LSTM input feature construction: progress feature: select the actual completion rate, planned completion rate, and progress deviation rate of each process in the past 14 days, a total of 3 dimensions; environmental disturbance feature: select rainfall, cumulative value, water level change, and extreme weather frequency such as heavy rain and gale in the past 14 days, a total of 3 dimensions; concatenate the two types of features into an LSTM input sequence, and set the time step to 14. Joint modeling and training: build a double-layer LSTM network: the first layer learns the time sequence rule of the progress feature, and the second layer integrates the environmental disturbance feature to output the future T period, such as T=7 days, construction progress prediction value and the influence coefficient of environmental disturbance on progress, such as heavy rain causing progress delay coefficient 0.8; loss function design: adopt progress prediction error + environmental disturbance influence error weighted sum, weight ratio 7:3, optimizer selects Adam, and the initial learning rate is set to 0.001.

[0055] Feature concatenation + attention mechanism fusion is adopted: concatenate the resource demand feature of each node in the future T period output by ST-GNN, with dimension of node number x resource type number, and the progress-environment joint feature output by LSTM, with dimension of 1 x influence coefficient; allocate feature weights through attention layer: for progress-sensitive resources such as concrete, progress delay needs to increase supply and give higher LSTM feature weight, and for environment-sensitive resources such as flood control equipment, heavy rain needs to increase scheduling and give higher ST-GNN spatial feature weight; form an initial demand prediction model, and output personnel, material, and equipment demand prediction values in each construction area in the future T period. Adopt training set, 70%-validation set, 20%-test set, and 10% data set division: verification index: resource demand prediction accuracy, MAE≤5% is qualified, progress correlation degree, and prediction demand and progress plan matching degree≥85% is qualified; super parameter adjustment: if the accuracy rate does not meet the standard, adjust the number of graph convolution layers of ST-GNN, 2-3 layers, and the number of hidden layer neurons of LSTM, 64-128, until the verification set index meets the requirements.

[0056] Step 103, adjust the inertia weight and cognitive factor of the demand prediction model by using IPSO improved particle swarm optimization algorithm, combine RL reinforcement learning to train the agent to learn the optimal scheduling strategy in the simulated environment, and obtain the target demand prediction model;

[0057] Specifically, in the embodiment, the particle swarm is initialized, the fitness of each particle is calculated, the particle position is adjusted according to the inertia weight, the cognitive factor and the social factor, a new parameter combination is generated, and when the global optimal fitness changes by less than or equal to 1% for 10 consecutive rounds, the optimal parameter combination is output. The state space is defined as a multi-dimensional vector of the current resource library, the demand prediction value of each region, the construction progress deviation and the environmental disturbance coefficient, and the action space is defined as a resource scheduling action; the DDPG deep deterministic policy gradient algorithm is used to construct the agent, the Actor network inputs the state and outputs a continuous scheduling action, the Critic network evaluates the value of the action, and the target demand prediction model is obtained.

[0058] Traditional PSO parameter defect analysis: the inertia weight of the traditional PSO controls the global search ability of the particle, and the cognitive factor controls the dependence of the particle on the optimal solution of itself, which are fixed values, which can easily lead to slow convergence speed or falling into local optimum in the later period, and need to be improved for the resource scheduling scene. IPSO parameter adaptive adjustment rule: inertia weight ω: adopt linear decrease + fitness feedback double mechanism, ω = 0.9 in the early stage of iteration, enhance global search, explore more resource scheduling strategies, ω linearly decreases to 0.4 with the iteration number in the later stage of iteration, and if the particle fitness, scheduling scheme satisfaction degree does not improve for 3 consecutive rounds, ω is temporarily increased by 0.1 to jump out of the local optimum; the cognitive factor c1 and the social factor c2: introduce nonlinear change, c1 = 2.5-0.01 x iteration number, depend on local optimum in early stage and depend on global optimum in later stage, c2 = 0.5+0.01 x iteration number, weak global guidance in early stage and strong global guidance in later stage.

[0059] Optimization target and particle coding: optimization target: minimize resource demand prediction error + scheduling cost, cost includes transportation cost, equipment idle cost and personnel standby cost; particle coding: encode the key parameters of the demand prediction model, such as the graph attention weight of ST-GNN and the learning rate of LSTM, as the particle dimension, and each particle corresponds to a set of model parameter combinations. IPSO iteration optimization process: initialize the particle swarm, the number of particles is 50, the dimension is 10-15, calculate the fitness of each particle, substitute the model to predict the demand, calculate the scheduling cost and the prediction error; update the particle speed and position: adjust the particle position according to the inertia weight, the cognitive factor and the social factor, the parameters corresponding to the global optimal particle, generate a new parameter combination; iteration termination condition: the iteration number reaches 100 rounds or the global optimal fitness changes by less than or equal to 1% for 10 consecutive rounds, output the optimal parameter combination, and update the initial demand prediction model.

[0060] 3 Scheduling scenario simulation environment building: State space S: defined as the current resource inventory, materials / equipment / personnel, demand forecast value of each area, construction progress deviation, environmental disturbance coefficient, multi-dimensional vector; Action space A: defined as resource scheduling actions, such as material transfer from A yard to B construction surface, quantity x tons, transportation time t hours, equipment transfer from C warehouse to D operation area, y units, scheduling priority p, personnel team E assigned to F process, z people, working time s hours; Reward function R: R = a x demand satisfaction rate - b x scheduling cost rate - g x progress delay rate, a = 0.5, b = 0.3, g = 0.2, reward when satisfaction rate ≥ 95%, penalty when cost overrun ≥ 10%. RL agent design: use deep deterministic policy gradient, DDPG algorithm to build agent: Actor network: input state S, output continuous scheduling action A, such as resource allocation quantity; Critic network: evaluate the value of action A, i.e. the expected reward R; Experience replay pool: store state-action-reward-next state samples, capacity 100,000, used for sample sampling during agent training.

[0061] Stage training strategy: Stage 1, IPSO pre-optimization: use IPSO optimized demand prediction model as the initial strategy of RL agent, agent learns scheduling actions based on the demand prediction results of the model in the simulation environment, such as adjusting the allocation quantity according to the predicted demand, train for 500 rounds, exploration rate ε = 0.8, high exploration; Stage 2, RL fine optimization: reduce exploration rate ε to 0.2, high utilization, agent iteratively learns in the simulation environment, and feedback the scheduling effect after each round of training, such as demand satisfaction rate and cost, to IPSO, which dynamically fine-tunes model parameters, such as adjusting inertia weight, forming a closed loop; Stage 3, convergence verification: when the average reward R of the agent is ≥ 0.8 for 100 consecutive rounds, the maximum score is 1.0, and the scheduling scheme stability is ≥ 90%, i.e. the difference between adjacent 5 rounds of scheduling actions is ≤ 5%, stop training.

[0062] Model output content: output demand prediction + scheduling strategy dual module: demand prediction module: future T period personnel, material, equipment demand quantity and confidence of each construction area, such as concrete demand 500 tons, confidence 92%; Scheduling strategy module: resource allocation priority, path, time, quantity suggestion. Model evaluation index: use prediction accuracy + scheduling efficiency dual dimension evaluation: prediction accuracy: resource demand prediction MAE ≤ 3%, progress correlation degree ≥ 90%; Scheduling efficiency: resource utilization rate ≥ 92%, scheduling cost reduction ≥ 8%, progress delay rate ≤ 3%, meet the index to determine the target demand prediction model.

[0063] Step 104, input the construction multi-source data into the target demand prediction model to output the scheduling scheme, and perform digital twin simulation verification and adjustment on the scheduling scheme to obtain the target scheduling scheme.

[0064] Specifically, in this embodiment, the simulation results are compared with the preset target, the problem root is traced back through the twin model, the adjustment scheme is formulated for the problem, and the target scheduling scheme is obtained.

[0065] Data format adaptation: convert the preprocessed multi-source data into a format recognizable by the target model: spatiotemporal graph data: input according to the node-edge structure required by ST-GNN, supplement the latest resource inventory such as daily stockyard inventory, environmental data such as future T-period weather forecast; time series data: update the latest deviation of construction progress, such as 80% of the process completion rate yesterday, equipment failure record, such as 2 excavators to be repaired; data verification: ensure the integrity of the input data, the missing rate is less than or equal to 1%, the timeliness, the data update time is less than or equal to 2 hours from the current.

[0066] Model output analysis: the target demand prediction model outputs the initial scheduling scheme, which specifically includes: personnel scheduling table: the allocation of each team, such as the steel reinforcement team and the pouring team, the working time, the task amount, such as the steel reinforcement team A is allocated to the dam body 3# area, and 50 tons of steel reinforcement is bound every day; material scheduling plan: material type, dispatching point / arrival point, quantity, transportation route, such as sand and stone from the west stockyard to the east pouring area, 1000 tons, along the shortcut 1, 3 times a day, transportation time window, such as 6:00-10:00 in the morning, avoiding the construction peak; equipment scheduling scheme: equipment type, dispatching warehouse / operation area, quantity, scheduling time, maintenance arrangement, such as 3 cranes from C warehouse to tunnel 2# operation area, to be in place before 14:00 today, and 1 hour of maintenance is arranged in the morning tomorrow; ④ cost and progress correlation table: the total cost corresponding to the scheduling scheme, transportation cost, labor cost, equipment rental cost, and the influence on each process progress, such as the scheme can ensure that the dam body pouring is advanced according to the plan.

[0067] Multi-dimensional twin model integration: geometric twin: based on BIM+GIS to build a 1:1 three-dimensional visualization model, including construction area terrain, buildings, dam body, tunnel, resource facilities, stockyard, equipment warehouse, transportation route; physical twin: integrate real-time data such as device running status and environmental parameters, simulate resource flow physical laws such as material transportation speed, load limit, and equipment operation efficiency curve; process twin: embed construction progress plan and resource scheduling rules, such as priority rules when material supply is insufficient, and environmental disturbance response mechanism, such as reducing transportation speed by 20% when it rains; ④ rule twin: import water conservancy and hydropower construction specifications, such as concrete pouring temperature requirements, equipment operation safety distance, and scheduling optimization objectives, such as resource utilization rate and cost control.

[0068] Multi-scenario simulation test: normal scenario simulation: run the twin model according to the initial scheduling scheme, simulate the resource scheduling process in the future T period, such as 7 days, record key indicators every hour: resource arrival rate, such as whether materials arrive on time at the construction site, resource utilization rate, such as equipment idle time, personnel standby time, progress completion rate, such as the actual progress of each process, cost consumption, such as whether the actual transportation cost exceeds the budget; disturbance scenario simulation: simulate common construction disturbances, such as heavy rain, equipment failure, material shortage, test the anti-interference ability of the scheduling scheme, such as simulating that the west material yard cannot supply materials due to heavy rain, observing whether the material supply is interrupted in the model, whether the standby scheme is triggered, such as enabling the east material yard. Real-time monitoring and data recording: real-time display of the scheduling process through the visualization interface of the twin model, such as resource transportation trajectory, construction site operation status, while automatically recording the index data of each round of simulation, generating simulation logs, such as Day110:00, the arrival rate of concrete in the east pouring area is 85%, due to traffic congestion of transportation vehicles, it is delayed for 30 minutes.

[0069] Simulation result analysis and bottleneck identification: indicator comparison: compare the simulation results with the preset targets, such as resource arrival rate ≥ 95%, resource utilization rate ≥ 92%, find out the unqualified items, such as the material arrival rate in the west pouring area is only 80% on Day2 due to traffic congestion on the transportation route; bottleneck positioning: trace back the problem source through the twin model, such as the congestion is caused by the intersection of shortcut 1 and equipment transportation route during peak hours; optimization adjustment: develop adjustment schemes for bottlenecks, such as optimizing the transportation route, changing the material transportation of the west pouring area to shortcut 2, adjusting the time window, changing the material transportation time from 8:00-12:00 to 5:00-9:00 in the morning, increasing standby resources, such as reserving 500 tons of emergency sand and gravel materials in the east material yard. Iterative verification and target scheme output: input the adjusted scheme into the twin model for simulation again, repeat the simulation-analysis-adjustment process until all key indicators meet the preset targets.

[0070] Its beneficial effect lies in that by fusing ST-GNN and LSTM model, the spatio-temporal dynamic characteristics of resource demand and the long-term dependence relationship of progress, environment and other time sequence disturbance factors can be captured at the same time, the accuracy of future resource demand prediction and management and scheduling efficiency are improved, and advanced and reliable data support is provided for scheduling decision, which changes passive response to active intervention. The IPSO algorithm is used to adaptively adjust the parameters, which greatly improves the convergence speed and accuracy of the demand prediction model; and the RL reinforcement learning is used to explore the optimal strategy in the simulation environment, effectively solving the complex resource scheduling optimization problem under multi-objective and multi-constraint, and generating the global optimal or approximate optimal solution with the lowest cost, the highest efficiency and the most balanced resources. The method embeds environmental disturbance factors in the modeling process, and verifies and dynamically adjusts the scheduling scheme through digital twinning technology, so that the system can quickly respond and generate a robust scheduling scheme when facing uncertain interference such as weather and geological mutations, and ensure the continuity and stability of construction.

[0071] Please refer to Figure 2 In the artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method, the resource demand of each construction area in the future T period is predicted based on the ST-GNN spatio-temporal graph neural network, and the LSTM long short-term memory network is combined to model the construction progress and environmental disturbance, and the demand prediction model includes the following steps,

[0072] Step 201, the construction area is divided into construction surface, resource supply point and transportation, and is defined as a graph node, the edges are constructed according to the association relationship between the nodes, and the time series data and spatial data of each node are embedded into the spatio-temporal graph as node features.

[0073] Step 202, the ST-GNN spatio-temporal graph neural network includes a spatial convolution and a time convolution double module, the spatial convolution layer captures the resource flow association between nodes through the GAT graph attention mechanism, and the time convolution layer captures the time trend through the 1D convolution.

[0074] The above describes the embodiment of the artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method, please refer to Figure 3 In the artificial intelligence-based water conservancy and hydropower construction resource management and scheduling system, the water conservancy and hydropower construction resource management and scheduling system includes the following modules,

[0075] The construction data acquisition module is used for acquiring Internet of Things data sensor data, BIM model data, GIS data and construction record data in the water conservancy and hydropower construction process, preprocessing the collected data, and obtaining construction multi-source data.

[0076] The prediction model establishment module is configured to predict resource demand of each construction area in a future T period based on an ST-GNN spatiotemporal graph neural network, and combine an LSTM long short-term memory network to jointly model construction progress and environmental disturbance, thereby obtaining a demand prediction model;

[0077] The prediction model optimization module is configured to adjust an inertia weight and a cognitive factor of the demand prediction model by using an IPSO improved particle swarm optimization algorithm, combine RL reinforcement learning to train an agent to learn an optimal scheduling strategy in a simulated environment, and thereby obtain a target demand prediction model.

[0078] The scheduling scheme generation module is configured to input construction multi-source data into the target demand prediction model to output a scheduling scheme, and perform digital twin simulation verification and adjustment on the scheduling scheme, thereby obtaining a target scheduling scheme.

[0079] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method is characterized by: The water conservancy and hydropower construction resource management and scheduling method comprises the following steps: Acquire IoT sensor data, BIM model data, GIS data, and construction record data during water conservancy and hydropower construction, pre-process the collected data, and obtain multi-source construction data; The resource demand of each construction area in the future T period is predicted based on the ST-GNN spatiotemporal graph neural network. The construction progress and environmental disturbance are jointly modeled by the LSTM long short-term memory network to obtain a demand forecasting model. The IPSO-modified particle swarm optimization algorithm is used to adjust the inertia weight and cognitive factor of the demand forecasting model. RL reinforcement learning is combined to train the intelligent agent to learn the optimal scheduling strategy in a simulated environment to obtain the target demand forecasting model. The multi-source construction data is input into the target demand forecasting model to output a scheduling plan, and the scheduling plan is verified and adjusted through digital twin simulation to obtain a target scheduling plan.

2. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method according to claim 1, characterized in that: The method of obtaining IoT data sensor data, BIM model data, GIS data, and construction record data during the water conservancy and hydropower construction process, and pre-processing the collected data to obtain multi-source construction data includes: Acquire IoT sensor data, BIM model data, GIS data, and construction record data during water conservancy and hydropower construction; Convert sensor binary data into JSON format, convert BIM model IFC format data into CSV structured data, convert GIS shp format data into GeoJSON format, and unify the timestamp format; Differentiated methods are used for missing value processing for different data types. Abnormal sensor data is eliminated using the 3σ principle, and abnormal records are filtered using business rules to obtain construction cleaning data. PCA principal component analysis is used to reduce the dimensionality of high-dimensional data in construction cleaning data to obtain construction multi-source data.

3. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method according to claim 1, characterized in that: The ST-GNN spatiotemporal graph neural network is used to predict the resource demand of each construction area within the future T period. The LSTM long short-term memory network is used to jointly model the construction progress and environmental disturbances to obtain a demand forecast model, including: The construction area is divided into construction areas, resource supply points, and transportation, which are defined as graph nodes. Edges are constructed based on the relationships between nodes. The time series data and spatial data of each node are used as node features and embedded in the spatiotemporal graph. The ST-GNN spatiotemporal graph neural network includes dual modules of spatial convolution and temporal convolution. The spatial convolution layer captures the resource flow association between nodes through the GAT graph attention mechanism, and the temporal convolution layer captures the timing trend through 1D convolution.

4. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method according to claim 1, characterized in that: The ST-GNN spatiotemporal graph neural network is used to predict the resource demand of each construction area within the future T period. The LSTM long short-term memory network is used to jointly model the construction progress and environmental disturbances to obtain a demand forecast model, including: A two-layer LSTM network is constructed. The first layer learns the temporal patterns of progress characteristics, and the second layer integrates environmental disturbance characteristics to output the predicted construction progress value for the future period and the impact coefficient of environmental disturbance on the progress. The resource demand characteristics of each node in the future period output by ST-GNN are spliced ​​with the progress and environment joint characteristics output by LSTM. Feature weights are assigned through the attention layer, giving high LSTM feature weights to progress-sensitive resources and high ST-GNN spatial feature weights to environment-sensitive resources.

5. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method according to claim 1, characterized in that: The IPSO improved particle swarm optimization algorithm is used to adjust the inertia weight and cognitive factor of the demand forecast model, and the RL reinforcement learning is combined to train the intelligent agent to learn the optimal scheduling strategy in a simulation environment to obtain the target demand forecast model, including: Initialize the particle swarm, calculate the fitness of each particle, adjust the particle position according to the inertia weight, cognitive factor, and social factor, and generate a new parameter combination. When the global optimal fitness changes by less than or equal to 1% for 10 consecutive rounds, output the optimal parameter combination.

6. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method according to claim 1, characterized in that: The IPSO improved particle swarm optimization algorithm is used to adjust the inertia weight and cognitive factor of the demand forecast model, and the RL reinforcement learning is combined to train the intelligent agent to learn the optimal scheduling strategy in a simulation environment to obtain the target demand forecast model, including: The state space is defined as a multidimensional vector of the current resource pool, demand forecast values ​​for each region, construction progress deviation, and environmental disturbance coefficient, and the action space is defined as resource scheduling actions; The DDPG deep deterministic policy gradient algorithm is used to build the intelligent agent. The Actor network inputs the state and outputs continuous scheduling actions. The Critic network evaluates the value of the action to obtain the target demand prediction model.

7. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling method according to claim 1, characterized in that: The step of inputting the multi-source construction data into the target demand forecasting model to output a scheduling plan, performing digital twin simulation verification and adjustment on the scheduling plan to obtain a target scheduling plan includes: Compare the simulation results with the preset targets, trace back to the root cause of the problem through the twin model, formulate an adjustment plan for the problem, and obtain the target scheduling plan.

8. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling system is characterized by: The water conservancy and hydropower construction resource management and scheduling system includes the following modules: The construction data acquisition module is used to obtain IoT data sensor data, BIM model data, GIS data, and construction record data during the water conservancy and hydropower construction process, and pre-process the collected data to obtain multi-source construction data; The prediction model building module is used to predict the resource demand of each construction area within the future T period based on the ST-GNN spatiotemporal graph neural network. It combines the LSTM long short-term memory network to jointly model the construction progress and environmental disturbances to obtain a demand prediction model. The forecast model optimization module is used to adjust the inertia weight and cognitive factor of the demand forecast model using the IPSO improved particle swarm optimization algorithm, and combine it with RL reinforcement learning to train the intelligent agent to learn the optimal scheduling strategy in a simulated environment to obtain the target demand forecast model; The scheduling plan generation module is used to input the multi-source construction data into the target demand prediction model to output a scheduling plan, perform digital twin simulation verification and adjustment on the scheduling plan, and obtain a target scheduling plan.

9. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling system according to claim 8, characterized in that: The prediction model optimization module includes the following submodules: Define submodules for defining the state space as a multidimensional vector of the current resource pool, demand forecast values ​​for each region, construction progress deviation, and environmental disturbance coefficients, and the action space as resource scheduling actions; Construct submodules for modules, which are used to build intelligent agents using the DDPG deep deterministic policy gradient algorithm. The Actor network inputs states and outputs continuous scheduling actions. The Critic network evaluates the value of actions to obtain the target demand prediction model.

10. The artificial intelligence-based water conservancy and hydropower construction resource management and scheduling system according to claim 8, characterized in that: The scheduling scheme generation module includes the following submodules: A sub-module is obtained, which is used to compare the simulation results with the preset targets, trace the root cause of the problem through the twin model, formulate an adjustment plan for the problem, and obtain the target scheduling plan.

Images