A container wharf yard dynamic security positioning method and system

Abstract

The application discloses a kind of based on multimodal environment perception and digital twin deduction container terminal yard dynamic security site selection method and system, belong to wisdom port and container terminal automation technical field.Method includes: deploying grid multi-modal environment sensor array in yard, real-time collection multidimensional environment parameter, output fusion environment parameter matrix;Build and continuously update yard digital twin three-dimensional model, and predict the environmental risk spatial distribution in future time window, generate yard environmental risk heat map;Extract the attribute data of incoming container to build coupling risk feature vector, and perform spatial matching and coupling calculation with environmental risk heat map, obtain the container-environment coupling risk index of each candidate idle slot;According to the preset hard constraint, divide the feasible set, solve the optimization model in the feasible set, and output the optimal safe slot.The safety satisfaction rate of the application is 100%, the pre-evacuation response time is shortened to 8 minutes, and the cargo loss in extreme weather is reduced by 60-80%.

Description

Technical Field

[0001] This application belongs to the field of smart port and container terminal automation technology, specifically involving a dynamic safety location method and system for container terminal yards based on multimodal environmental perception and digital twin simulation. Background Technology

[0002] With the continued growth of global trade, container terminals, as a key node in the modern logistics system, are receiving increasing attention for their operational safety and efficiency. Yard Location Allocation (YLA) is a core component of terminal operation planning, and the quality of its decisions directly impacts yard utilization, operational efficiency, and the level of safety risk control.

[0003] In recent years, with the rapid development of artificial intelligence and automation technologies, a large amount of research has focused on using deep reinforcement learning, heuristic optimization algorithms and digital twin technology to improve the efficiency and intelligence of location selection decisions.

[0004] However, existing container terminal yard location technologies generally prioritize efficiency over safety, lack deeply integrated systematic solutions, and cannot meet the actual needs of dynamic and safe container location in yards under extreme weather and complex environmental conditions. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and provide a method and system for dynamic safety location selection of container terminal yards based on multimodal environmental perception and digital twin inference, so as to solve the defects of the prior art that only focuses on operational efficiency optimization while ignoring environmental safety constraints and lacks a refined coupling assessment of container attributes and environmental risk factors.

[0006] To achieve the above objectives, this application adopts the following technical solution: This application provides a dynamic safety location selection method for container terminal yards based on multimodal environmental perception and digital twin simulation, including the following steps: Step S1: Deploy a gridded multimodal environmental sensor array in the stockyard to collect multidimensional environmental parameters at a preset period, and output a fused environmental parameter matrix after data fusion and preprocessing; Step S2: Build and continuously update the digital twin 3D model of the stockpile; Step S3: Based on the fused environmental parameter matrix, run the environmental risk simulation model in the digital twin engine to predict the spatial distribution of environmental risks within the future time window and generate a heat map of environmental risks in the storage yard. Step S4: When a container arrives at the site to be selected, extract the attribute data of the container and construct a coupled risk feature vector; Step S5: Spatial matching and coupling calculation are performed between the coupled risk feature vector and the environmental risk heat map to obtain the container-environment coupling risk index CRI(i) for each empty slot in the yard; Step S6: Using CRI less than a preset safety threshold τ as a hard constraint, divide all candidate slots into a feasible set S_feasible={i | CRI(i)<τ} and an infeasible set; if the feasible set is not empty, solve the optimization model with the goal of minimizing the operation cost within the feasible set and output the optimal slot; if the feasible set is empty, trigger an alarm and enter the pre-evacuation process.

[0007] Preferably, the method further includes: Step S7: Compare the fused environmental parameter matrix with the preset secondary warning threshold in real time. When any environmental parameter exceeds the secondary warning threshold, the adaptive pre-evacuation process is automatically triggered to track the risk evolution path in the digital twin and identify the high-risk container set whose CRI exceeds the evacuation threshold τ_evac. Step S8: With the goal of minimizing the total evacuation cost, solve the redistribution scheme for the high-risk container set within the safe area, generate an evacuation instruction sequence, and issue it for execution.

[0008] Preferably, in step S7, the secondary warning threshold includes: 10-minute average wind speed ≥ 17.2 m / s; hourly rainfall ≥ 30 mm; tide height exceeding 80% of the wharf's design elevation; peak ground acceleration ≥ 0.05 g; and / or, the evacuation threshold τ_evac is dynamically set according to the current safety level: evacuation threshold τ_evac = 0.5 for the first safety level, 0.7 for the second safety level, and 0.85 for the third safety level, and τ_evac is always higher than the safety threshold τ of the current safety level; And / or, in step S8, the objective function of the total evacuation cost is min Σ (w_i×d_i), where w_i is the evacuation weight of the i-th high-risk container, and d_i is the horizontal transport distance of the container from its original storage location to the newly assigned safe storage location; and / or, the constraints of the redistribution scheme include: the newly assigned storage location is located within the safe area of ​​CRI<τ in the current environmental risk heat map, and the evacuation path does not pass through waterlogged areas or high-risk areas; and / or, the evacuation instruction sequence is generated in priority order, with the priority order being: hazardous chemical containers first, loaded containers second, and empty containers last.

[0009] Preferably, the multimodal environmental sensor array in step S1 is deployed as an N×M regular grid with a grid spacing of 50-100 meters; and / or, the multidimensional environmental parameters include: wind speed and direction at a height of 10 meters, tide height, rainfall, atmospheric visibility, and peak ground acceleration; and / or, the data fusion and preprocessing include: timestamping all sensor data using the Network Time Protocol, using the 3σ criterion for outlier detection and filtering, using the Kalman filter algorithm for smoothing, and outputting an environmental parameter vector E(t)={v(t), θ(t), h(t), r(t), d(t), a(t)}, where v is wind speed, θ is wind direction, h is instantaneous tide height, r is hourly rainfall, d is atmospheric visibility, and a is peak ground acceleration.

[0010] Preferably, in step S2, the digital twin 3D model includes a terrain elevation model, a real-time container stacking status model, a drainage network topology model, and a windproof structure shielding topology model. The terrain elevation model generates a digital elevation matrix of the entire storage yard based on LiDAR point cloud data or engineering survey data, with a spatial resolution of 0.5m × 0.5m. The real-time container stacking status model obtains the container number, size, weight, hazardous chemical class, current slot coordinates and stacking layer of each container in the yard in real time through the yard crane PLC data interface and OCR container number recognition system, and updates the stacking status in the digital twin in real time in a three-dimensional voxel manner. The drainage network topology model is a directed graph topology of the drainage network constructed based on the design drawings of the yard drainage project. The nodes include rainwater inlets, inspection wells, drainage pipe sections and outlets, and the edges are labeled with pipe diameter, slope and Manning roughness coefficient. The windproof structure obstruction topology model uses a ray tracing algorithm to pre-calculate the equivalent windward area reduction factor for each storage location in the yard under different wind directions due to obstruction from surrounding container stacks and buildings.

[0011] Preferably, the environmental risk simulation model in step S3 includes: A lightweight CFD wind field model, based on a simplified steady-state solver of the RANS equations, uses container stacks as the source of porous medium resistance terms, and outputs the wind speed distribution field at a height of 10 meters above the horizontal plane of the storage yard. The flooding simulation model based on SCS-CN uses real-time rainfall as input and combines topographic elevation model and drainage network topology to predict the water depth of each tank in the storage yard within the future time window ΔT. The seismic response spectrum extrapolation model, based on the site response spectrum and the dynamic characteristics of container stacking, calculates the dynamic amplification factor and overturning risk of containers under different stacking layers; The environmental risk heat map H(x,y,t+ΔT) is generated by fusing the outputs of the above three extrapolation models after normalization, where (x,y) are the planar coordinates of the stockpile, and H takes a continuous value between 0 and 1.

[0012] Preferably, the coupled risk feature vector R_c={m, S_wind, class_haz, tier,z} mentioned in step S4, where m is the container mass, S_wind is the equivalent windward area, class_haz is the hazardous chemical risk level, tier is the target stacking layer number, and z is the bottom elevation of the tank. And / or, the formula for calculating the container-environment coupling risk index CRI(i) in step S5 is: CRI(i)=α·f_wind(i)+β·f_flood(i)+γ·f_seismic(i)+δ·f_hazmat(i) Wherein, f_wind is the wind-induced overturning risk component, calculated as (equivalent windward area × wind speed squared × stacking height) / (mass × base width × gravitational acceleration); f_flood is the water accumulation risk component, calculated as the normalized value of the difference between the predicted water accumulation depth and the bottom elevation of the tank; f_seismic is the seismic risk component, calculated as the predicted peak ground acceleration × number of stacking layers / seismic design acceleration; f_hazmat is the hazardous chemical risk amplification factor, with a value of hazardous chemical level / 9; α, β, γ, and δ are weighting coefficients, satisfying α + β + γ + δ = 1.

[0013] Preferably, the weighting coefficients α, β, γ, and δ are dynamically switched according to the current warning level: During typhoon warning periods: α=0.5, β=0.2, γ=0.1, δ=0.2; During a rainstorm warning: α=0.1, β=0.6, γ=0.1, δ=0.2; During normal operation: α=0.2, β=0.2, γ=0.1, δ=0.5.

[0014] Preferably, the preset safety threshold τ in step S6 is set according to the current terminal safety level: Level 1 safety level τ=0.3, Level 2 safety level τ=0.5, Level 3 safety level τ=0.7; and / or, the objective function of the optimization model is min C=ω_1×D(i)+ω_2×F(i), where D(i) is the horizontal transport distance of the candidate container from the gate or ship berth to the tank i, F(i) is the cost of reloading caused by being assigned to tank i, and ω_1 and ω_2 are weighting coefficients.

[0015] Secondly, this application provides a dynamic safety location system for container terminal yards based on multimodal environmental perception and digital twin simulation, comprising: The multimodal environment sensing layer is used to deploy a gridded multimodal environment sensor array in the yard to collect multidimensional environmental parameters at a preset period. The data fusion and preprocessing layer is used to receive the raw sensor data from the multimodal environment perception layer, perform timestamp synchronization, outlier filtering and Kalman filtering smoothing, and output a fused environment parameter matrix with a unified time resolution. The digital twin simulation engine is used to build and maintain a high-precision three-dimensional digital twin model of the storage yard, and to infer the spatial distribution of environmental risks within a future time window based on the fused environmental parameter matrix, generating a heat map of environmental risks in the storage yard. The container-environment coupling risk assessment module is used to extract attribute data for each candidate container to construct a coupling risk feature vector, and perform spatial matching and coupling calculation with the environmental risk heat map to obtain the container-environment coupling risk index CRI(i) for each candidate slot. The dynamic safety location selection decision module is used to construct a feasible set with CRI less than a preset safety threshold τ as a hard constraint, solve an optimization model within the feasible set with the goal of minimizing the operating cost, and output the optimal slot. An adaptive pre-evacuation rescheduling module is used to trigger an alarm and switch to the pre-evacuation process if the feasible set is empty.

[0016] Preferably, the adaptive pre-evacuation rescheduling module is also used to automatically trigger the pre-evacuation process when environmental parameters exceed the preset secondary warning threshold, identify high-risk container sets, and generate a redistribution scheme with the goal of minimizing the total evacuation cost.

[0017] Preferably, in the multimodal environmental perception layer, the sensors deployed in each grid node include: ultrasonic anemometer, tide gauge, rain gauge, visibility sensor and triaxial accelerometer seismic sensor; each sensor uploads data in real time through industrial Ethernet or LoRa wireless sensor network, with a data acquisition cycle of 5-10 minutes.

[0018] Preferably, the dynamic safety location selection decision module uses the ε-constraint method to construct an optimization model, with CRI less than a preset safety threshold τ as an inviolable hard constraint, and minimizing the weighted sum of horizontal transport distance and container handling cost as the optimization objective. The optimal slot is solved by an integer programming solver or a heuristic algorithm. The decision results are sent to the yard crane PLC and truck scheduling system for execution via MQTT or OPC UA protocol.

[0019] Preferably, in the adaptive pre-evacuation rescheduling module: the secondary early warning thresholds include wind speed threshold, rainfall threshold, tide height threshold, and earthquake acceleration threshold; the high-risk container set is the set of containers whose CRI exceeds the evacuation threshold τ_evac; the reassignment scheme takes minimizing the total evacuation cost as the objective function and constrains the newly assigned slots to be located within the safe area where CRI < τ; the system generates an evacuation instruction sequence according to the priority order of hazardous chemical containers first, loaded containers second, and empty containers last.

[0020] Compared to existing technologies, the beneficial effects achieved by this application may include one or more of the following aspects: (1) For the first time, multimodal environmental perception grid data and digital twin simulation are deeply integrated to realize a paradigm upgrade from passive alarm to active safety location selection. Existing weather warning schemes (such as CN118863154A) can only generate alarms and suspend operations, and cannot perform safe relocation at the tank level based on warning information. However, this invention deploys a gridded multimodal environmental sensor array in the yard to collect multi-dimensional environmental parameters such as wind speed, wind direction, tides, rainfall, visibility, and seismic acceleration in real time. These parameters are used as inputs to a digital twin simulation engine to predict the spatial distribution of environmental risks within future time windows. This allows for proactive avoidance of high-risk areas during the container arrival and relocation stage, achieving proactive and safe relocation and significantly improving the terminal's risk response capabilities under extreme weather conditions.

[0021] (2) The Container-Environment Coupled Risk Index (CRI) model enables differentiated and refined safety assessments, avoiding a one-size-fits-all approach. Existing site selection schemes lack a quantitative coupling relationship between container attributes and environmental risk factors, leading to a one-size-fits-all approach to safety measures. This results in either overly conservative measures that negatively impact operational efficiency or safety hazards. The CRI model proposed in this invention quantitatively couples container attributes (mass, equivalent windward area, hazardous chemical risk level, target stacking layer number, bottom elevation) with real-time environmental conditions (wind speed, tide height, rainfall, seismic acceleration). By calculating the coupling risk index for each candidate slot using a weighted multi-factor model, it enables differentiated safety assessments based on container type. This differentiated assessment mechanism effectively solves the problem of one-size-fits-all safety measures, maximizing operational efficiency while ensuring a safety baseline.

[0022] (3) The adaptive pre-evacuation mechanism automatically completes the preventive transfer of high-risk containers before extreme weather arrives, significantly reducing the risk of cargo damage and safety accidents. Existing technologies lack adaptive pre-evacuation capabilities based on environmental threshold triggers. Even after typhoon or heavy rain warnings are issued, evacuation plans still require manual formulation, resulting in slow response times and a high risk of overlooking high-risk containers. The environmental threshold triggering mechanism established in this invention automatically triggers the pre-evacuation process when environmental parameters exceed the secondary warning threshold or when the feasible location set is empty. By rapidly extrapolating risk spread paths in a digital twin, it accurately identifies high-risk container sets, then automatically solves a redistribution scheme with the objective of minimizing the total evacuation cost, and generates and executes a sequence of evacuation instructions according to priority. This mechanism reduces the pre-evacuation response time from hours in existing technologies to minutes, significantly reducing cargo damage and safety risks under extreme weather conditions.

[0023] (4) The ε-constraint framework of hard safety constraints and soft efficiency optimization ensures both safety baseline and operational efficiency. Existing site selection schemes primarily optimize operational efficiency without considering safety constraints as hard constraints, posing safety risks under extreme weather conditions. This invention employs the ε-constraint method to construct a site selection optimization model, treating the CRI safety threshold τ as an inviolable hard constraint (i.e., only slots with CRI < τ are included in the feasible set), and optimizing operational efficiency (transportation distance, handling costs) as a soft objective. This approach minimizes efficiency losses caused by safety constraints while ensuring a safety baseline.

[0024] (5) The system architecture is modular and open, with good scalability and maintainability. This invention divides the system into six functional modules (multimodal environment perception layer, data fusion and preprocessing layer, digital twin inference engine, container-environment coupling risk assessment module, dynamic safety location decision module, and adaptive pre-evacuation rescheduling module). The modules communicate with each other through standard data interfaces (such as OPC UA, MQTT, REST API). The internal implementation details of the modules are transparent to the outside world, which facilitates subsequent functional expansion and algorithm upgrades.

[0025] (6) Experimental verification has shown that the present invention has produced unexpected and significant technological progress. An experiment based on actual data from a coastal container terminal shows that: (a) Safety satisfaction rate: increased from 45% of existing technology to 100% (+55 percentage points); (b) Pre-evacuation response time: reduced from hours (average 95 minutes) for manual decision-making to 8 minutes (-92%). (c) Cargo loss due to extreme weather: reduced by approximately 60-80%; (d) Loss of operational efficiency: only 12-18% (compared to pure efficiency optimization methods); (e) Turnover rate: reduced from 2.5% to 0.8%; (f) Annual reduction in cargo loss per terminal: approximately RMB 3-5 million; (g) Annual reduction in accident losses per wharf: approximately RMB 1-2 million; (h) Investment payback period: 2.5-5.5 years. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall architecture of the container terminal yard dynamic safety positioning system based on multimodal environmental perception and digital twin simulation of the present invention. Figure 2 This is a flowchart illustrating the dynamic safety location method for container terminal yards based on multimodal environmental perception and digital twin simulation, as described in this invention. Figure 3 This is a schematic diagram of the structure of the digital twin inference engine of the present invention; Figure 4 This is a schematic diagram of the sensor deployment in the multimodal environment perception layer of the present invention; Figure 5 This is a schematic diagram of the container-environment coupling risk index (CRI) calculation process of the present invention; Figure 6 This is a schematic diagram of the optimization solution process for the dynamic safety location selection decision module of the present invention; Figure 7 This is a schematic diagram of the workflow of the adaptive pre-evacuation rescheduling module of the present invention. Detailed Implementation

[0027] The specific embodiments of the technical solution of this application will be described in detail below with reference to the accompanying drawings. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0028] Existing technical solutions generally have the following shortcomings: On the one hand, existing container allocation schemes primarily focus on optimizing operational efficiency, failing to consider real-time environmental safety risks as a core constraint in location selection decisions. Analysis reveals that the root cause lies in traditional terminal operation management treating safety and efficiency as two independent dimensions, assuming that safety measures inevitably sacrifice efficiency. However, under extreme weather conditions, safety incidents such as container overturning, flooding, and earthquake collapses can lead to complete operational disruptions, with the efficiency loss far exceeding the small increase in transport distance due to safety constraints.

[0029] For example, Shanghai International Port Group (CN119047691A) uses the deep reinforcement learning algorithm D3QN to train a Markov decision model, enabling real-time container location assignment, effectively reducing the number of container handling operations and improving loading efficiency. However, this solution only optimizes operational efficiency and does not consider the impact of environmental factors such as wind speed, rainfall, and earthquakes on container storage safety. Similarly, Guangzhou Port Group (CN121211960B) constructs a smart container receiving gravity model based on the UCT-MCTS framework, optimizing container location by maximizing container attraction and minimizing repulsion, thus improving port operational efficiency. However, this gravity model does not incorporate environmental risk factors into its gravity calculation framework, making it unable to guarantee yard safety during typhoons or heavy rain. For example, the container location allocation method based on a large model-driven intelligent agent network, jointly applied for by Tianjin Port Second Container Terminal Co., Ltd. and Tianjin University (CN121860265A), employs a dual-stream attention mechanism for multi-view hybrid encoding of containers and uses a large model for structured evaluation and intelligent sorting of containers, significantly improving container loading efficiency. However, this solution also fails to consider environmental safety constraints as a core dimension of location selection decisions, posing serious safety hazards under extreme weather conditions. Furthermore, the container layout optimization method for port yards proposed by Bayes Computing (Tianjin) Information Technology Co., Ltd. (CN121031893A) calculates vectorized features using a target weight calculation model to screen target locations, achieving optimized layout based on global state. However, the weight calculation of this method does not incorporate real-time environmental risk factors, making it unable to perform differentiated safe location selection based on dynamically changing weather conditions.

[0030] After repeated research, this application proposes an ε-constraint framework of "hard safety constraints + soft efficiency optimization", which takes safety as an insurmountable bottom line and optimizes efficiency within the safe feasible domain, thereby fundamentally solving the problem of the conflict between safety and efficiency.

[0031] Secondly, existing digital twin-based terminal solutions focus on yard state mapping and gravitational field simulation, without incorporating real-time fusion and simulation of multimodal environmental sensor grid data with the digital twin. In other words, existing digital twin systems mainly rely on historical data and static models for storage state simulation, failing to integrate real-time environmental perception data, resulting in a disconnect between digital twins and environmental perception.

[0032] For example, the digital twin simulation method for container terminal yard management applied for by Xiamen Container Terminals Group Co., Ltd. (CN119294239A) generates the initial state of the digital twin platform by acquiring current production data related to container terminal and yard management, sets the probability distribution of random events based on historical operation patterns, and selects the location of incoming containers based on the yard management gravitational field model. However, the digital twin model of this scheme is mainly used to simulate the yard operation state, and does not integrate real-time data from gridded multimodal environmental sensors. It cannot extrapolate and predict wind field distribution, flooding range, and seismic response within future time windows, and therefore cannot achieve proactive and safe location selection based on environmental risk prediction. The related patent of Shanghai Maritime University (CN114595502A) has been withdrawn, further illustrating the technical defects of existing digital twin schemes in environmental risk integration and extrapolation. The "Ship Bay Position Optimization and Positioning Method Based on Digital Twin" (CN117788568B), jointly applied for by Tianjin Port Pacific International Container Terminal Co., Ltd. and Shanghai Xinwei Intelligent Technology Co., Ltd., achieves real-time positioning of the hook to the bay position through a digital twin model. It mainly solves the positioning accuracy problem during the loading and unloading process of ships, but does not involve the perception of yard environmental risks or the safe positioning of containers.

[0033] In-depth analysis revealed that the environmental risks of storage yards exhibit significant spatiotemporal dynamics—the risk difference for the same storage location under different wind speeds and rainfall intensities can reach several times or even tens of times. If real-time environmental parameters are not used as input for the digital twin, the inference results will deviate significantly from physical reality. Therefore, this application proposes a three-layer architecture: a multimodal environmental perception layer → a data fusion and preprocessing layer → a digital twin inference engine. Kalman filtering is used to achieve spatiotemporal synchronization and smooth fusion of multi-source heterogeneous sensor data. The fused environmental parameter matrix is ​​then used as the real-time input to the digital twin inference engine, thus realizing a qualitative leap in digital twins from "static mapping" to "dynamic inference."

[0034] In three aspects, existing technologies adopt uniform safety standards for different types of containers, failing to fully consider the impact of the container's own attributes on risks. For example, existing weather warning patents only generate alarms and suspend operations, without involving dynamic safety repositioning of containers at the slot level, and lack differentiated safety assessments.

[0035] For example, a terminal early warning system and method based on a meteorological prediction model (CN118863154A) collects data such as temperature, humidity, wind speed, wind direction and rainfall in real time by deploying multiple meteorological sensors in the terminal area. It uses a dual attention mechanism long short-term memory neural network (DA-LSTM) to make meteorological predictions and generates strong wind warnings and rainstorm warnings based on the prediction results. The production scheduling module adjusts the work plan according to the warning information (such as suspending high-altitude operations and hoisting operations). However, this solution can only generate warning information and suspend related operations. It cannot perform safe relocation of containers already stacked in the yard and those waiting to enter the yard according to the warning level. This results in the need for manual intervention to relocate containers after the warning is triggered, which is slow and prone to missing high-risk containers.

[0036] Analysis of extensive experimental and historical accident data revealed that: empty containers face more than 10 times the risk of overturning at the same wind speed compared to loaded containers; hazardous chemical containers face significantly higher risks of secondary disasters after being submerged in water compared to ordinary containers; and the wind-induced overturning moment approximately doubles with each additional stacking layer. Based on these findings, this application constructs a Container-Environment Coupled Risk Index (CRI) model, which quantitatively couples five attributes—container mass, equivalent windward area, hazardous chemical risk level, target stacking layer number, and bottom elevation—with four environmental conditions—wind speed, tide height, rainfall, and seismic acceleration—to achieve differentiated and refined safety assessments.

[0037] Fourthly, existing technologies lack a quantitative model for the coupling risk between containers and the environment, making it impossible to perform differentiated safety positioning based on the coupling relationship between the container's own attributes and real-time environmental conditions. Different containers have significant differences in attributes such as weight, size, hazardous chemical class, windward area, and number of stacking layers, and their safety risks vary even under the same environmental conditions.

[0038] For example, hazardous chemical containers face significantly higher risks in strong winds or high temperatures than ordinary dry cargo containers, and empty containers are at much higher risk of overturning in strong winds than loaded containers. Existing location selection schemes lack a quantitative coupling relationship between container attributes and environmental risk factors, failing to provide differentiated safety assessments and location decisions for different types of containers. This results in a one-size-fits-all approach to safety measures, either being overly conservative and impacting operational efficiency, or posing safety hazards. The container storage location selection method based on stacking behavior characteristics (CN202411085076.9), jointly developed by Chongqing Normal University and Chongqing Guoyuan Container Terminal Co., Ltd., uses neural networks to predict container stacking time and combines this prediction with a rule-based storage location selection method to apply the results to container area and location selection, effectively reducing the container turnover rate. However, this method does not consider the coupling relationship between container attributes and environmental risks, thus failing to achieve differentiated safety location selection.

[0039] By combining the container-environment coupling risk index (CRI) model constructed above, differentiated and refined safety assessment and site selection were achieved.

[0040] In five aspects, existing early warning systems can only generate alarm information and cannot automatically translate it into specific evacuation action plans; that is, they lack an adaptive pre-evacuation mechanism triggered by environmental thresholds. In particular, before extreme weather events such as typhoons and heavy rains, terminals need to promptly transfer containers from high-risk areas within the yard to safe areas to reduce cargo damage and safety risks. Existing technical solutions lack an adaptive pre-evacuation mechanism triggered by environmental thresholds, and cannot automatically generate and execute the optimal evacuation plan after identifying high-risk container sets through digital twin simulations.

[0041] For example, the intelligent container terminal location calculation method used by Qingdao Port International Co., Ltd. and other companies reduces congestion or excessive dispersion caused by automatic container allocation through computer programs, but it does not involve a pre-evacuation and rescheduling mechanism based on environmental risk prediction. The container unloading stacking planning method (CN120494297A) applied for by CCCC First Harbor Engineering Survey and Design Institute Co., Ltd. achieves unloading stacking planning through cluster analysis and distance matrix calculation, which improves unloading efficiency, but it also lacks environmental adaptive pre-evacuation capabilities.

[0042] Through in-depth analysis, it was recognized that the pre-evacuation window before extreme weather events is extremely limited (usually only a few hours). Manually formulating evacuation plans is too time-consuming and prone to overlooking high-risk containers. To address this, this application establishes a pre-evacuation chain that includes "automatic environmental threshold triggering → digital twin risk simulation → automatic identification of high-risk containers → automatic solution of evacuation plans → automatic issuance of instructions," reducing response time from hours to minutes.

[0043] Further exploration revealed the following technological gaps in the existing technology in this field: (1) There is a lack of technical solutions to deeply integrate multimodal environmental perception data with digital twin simulation. Existing digital twin systems are only used for the visualization and mapping of yard status, and do not use real-time environmental parameters as input to the simulation engine, so they cannot predict future risks. (2) There is a lack of a quantitative assessment model for container-environment coupling risks. Existing location selection methods have not established a quantitative coupling relationship between container attributes (mass, windward area, hazardous chemical grade, etc.) and environmental risk factors (wind speed, tide height, rainfall, etc.), resulting in a "one-size-fits-all" approach to safety measures and an inability to achieve differentiated and refined assessment. (3) There is a lack of pre-evacuation mechanism that is automatically triggered based on environmental thresholds. The existing early warning system can only generate alarm information and cannot automatically identify high-risk container sets and generate the optimal evacuation plan. The reliance on manual decision-making leads to slow response speed and easy omission of high-risk containers.

[0044] Therefore, existing container terminal yard location technologies generally prioritize efficiency over safety, lacking a systematic solution that deeply integrates multimodal environmental perception, digital twin simulation, and container-environment coupled risk quantification models. This fails to meet the actual needs of dynamic and safe container location selection in extreme weather and complex environmental conditions. There is an urgent need in this field for an intelligent location selection method and system that can perceive multidimensional environmental risks in real time, predict future environmental risk distribution through digital twin simulation, perform differentiated safe location selection based on coupled risk quantification models, and have the capability to automatically trigger pre-evacuation based on environmental thresholds.

[0045] Further in-depth analysis of existing technologies has revealed two fundamental flaws in current container terminal yard location technologies: Flaw 1: The separation of safety and efficiency. Existing location selection schemes treat safety and efficiency as two independent dimensions, typically by adding a safety penalty term to the efficiency optimization objective function (i.e., safety as a soft constraint). However, under extreme weather conditions, this approach cannot guarantee a safety baseline, especially when the efficiency gains outweigh the safety penalties; the system will still allocate containers to high-risk areas. More seriously, in the event of a safety accident (such as container overturning or flooding), operations will be completely halted, and efficiency will drop to zero. Therefore, safety must be an insurmountable hard constraint, not a trade-off soft objective.

[0046] Defect 2: The disconnect between static models and dynamic risks. Existing digital twin systems are mainly used for visual mapping of storage conditions and gravitational field simulation. Their models are static (based on historical data and fixed parameters), while the environmental risks of storage yards are highly dynamic (wind speed changes with typhoon paths, water accumulation changes with rainfall intensity). Static models cannot reflect the spatiotemporal dynamics of risks, and therefore cannot achieve proactive site selection based on environmental risk prediction.

[0047] To address the two major shortcomings mentioned above, this application proposes a three-step, systematic solution: Step 1: Upgrade environmental risks from "external warnings" to "internal constraints," making risks perceptible. Existing weather warning systems only generate warning information at the terminal level, failing to map environmental risks to every storage location in the yard. It is essential to use a gridded sensor array to perceive multi-dimensional environmental parameters (wind speed, wind direction, tide height, rainfall, visibility, seismic acceleration) across the entire storage yard in real time. After data fusion, a unified environmental parameter matrix is ​​formed, transforming environmental risks from "macro-level warnings" to "micro-level measurability."

[0048] The second step is to upgrade the digital twin from a "static mapping" to a "dynamic simulation" to achieve predictable risks. Existing digital twins can only reflect the current storage status and cannot predict future risks. It is necessary to use the real-time environmental parameter matrix as input to the digital twin simulation engine. Through CFD wind field simulation, flooding simulation, and seismic response spectrum simulation, the environmental risk distribution of each storage location within a future time window can be predicted, generating an environmental risk heat map, thus transforming risk from "currently visible" to "future predictable."

[0049] The third step is to "physically couple" container attributes with environmental risks to achieve differentiated risk assessment. Different types of containers have vastly different risks under the same environmental conditions (the risk of an empty container tipping over is more than 10 times that of a loaded container). Therefore, it is necessary to establish a quantitative coupling model between container attributes (mass, windward area, hazardous chemical class, number of stacking layers, bottom elevation) and environmental risk factors, and calculate the Coupled Risk Index (CRI) for each slot for a specific container, thus achieving differentiated safety assessment from "uniform standards" to "one policy per container".

[0050] Based on the above ideas, the present invention employs one or more of the following core technical means to constitute the corresponding technical solution.

[0051] (i) Deep integration of multimodal environment perception and digital twins to solve the connection problem between "perception and inference". The core technical feature is: reference Figure 4 The diagram illustrates the deployment of an N×M regular grid of multimodal sensors (grid spacing 50-100 meters) within the storage yard. Each sensor uploads data via industrial Ethernet or LoRa wireless sensor network. The data fusion layer performs timestamp synchronization (NTP protocol), outlier filtering (3σ criterion), and Kalman filtering smoothing on the multi-source heterogeneous sensor data, outputting an environmental parameter matrix E(t) with a unified time resolution. This matrix is ​​pushed to the digital twin simulation engine in real time via shared memory or message queue, transforming the digital twin from a "historical data-driven static model" into a "real-time environmental data-driven dynamic simulation engine." Figure 4 The image shows an N×M sensor grid layout (spacing 50-100 meters) and five types of sensors equipped in each grid node: ultrasonic anemometer, laser rain gauge, radar tide gauge, visibility meter, and MEMS seismic accelerometer.

[0052] It is important to emphasize that the "fusion" in this application is not a simple data superposition, but rather the realization of spatiotemporal synchronization and noise suppression of multi-source sensor data. For example, data fusion processing is performed through Kalman filtering, so that heterogeneous data with different sampling frequencies (1Hz for anemometers, 100Hz for seismographs, and 1 time / hour for rain gauges) can work together on a unified time reference.

[0053] (II) Parallel operation of the three major projection models and generation of heat maps to solve the projection problem of "current situation → prediction". The core technological features are: a digital twin simulation engine that runs three lightweight physical models in parallel—a CFD wind field model based on RANS equations (using container stacks as the source of porous media resistance, with a solution time of <30 seconds), a flooding model based on the SCS-CN runoff curve number method (using real-time rainfall as input, combined with DEM and drainage network topology), and a seismic response spectrum model (based on site response spectrum and stacked dynamic characteristics). The outputs of the three models are normalized and then fused into an environmental risk heatmap H(x,y,t+ΔT), where H takes continuous values ​​between 0 and 1.

[0054] The key innovations of this technology lie in "lightweighting" and "parallelization": traditional CFD calculations take hours to days, while this invention reduces the calculation time to less than 30 seconds by simplifying RANS equations and using the porous media assumption, enabling environmental risk prediction to be integrated into the real-time operation process of the terminal, rather than just for offline simulation analysis.

[0055] (III) The Container-Environment Coupled Risk Index (CRI) model addresses the quantitative problem of "unified standards → differentiated assessment". The core technical feature lies in constructing a coupled risk feature vector R_c containing five dimensions of container attributes (mass m, equivalent windward area S_wind, hazardous chemical class_haz, target stacking layer number tier, bottom elevation z), and quantitatively coupling it with four-dimensional environmental risk factors (wind speed v, water depth d_flood, seismic acceleration a, hazardous chemical amplification factor class_haz / 9). The calculation formula is as follows: CRI(i)=α·f_wind(i)+β·f_flood(i)+γ·f_seismic(i)+δ·f_hazmat(i); The weighting coefficients are dynamically switched according to the warning level (for example, typhoon: α=0.5, β=0.2, γ=0.1, δ=0.2; rainstorm: α=0.1, β=0.6, γ=0.1, δ=0.2; daily: α=0.2, β=0.2, γ=0.1, δ=0.5).

[0056] The key innovation of this technology lies in the physical meaning of the CRI calculation formula, as follows: f_wind = (S_wind×v²×h_tier) / (m×w_base×g) is essentially the ratio of wind-induced overturning moment to the container's gravity moment, with a clear physical meaning (a ratio > 1 indicates an extremely high risk of overturning); f_flood = max(0, d_flood-z) / H_max reflects the safety margin between the tank elevation and the predicted water depth; f_seismic = a×tier / a_max reflects the amplification effect of the number of stacked layers on the seismic response. These formulas are not empirical fits, but are derived based on the fundamental principles of wind engineering, hydraulics, and structural dynamics.

[0057] (iv) ε-constrained safety hard constraint optimization framework to solve the decision problem of "safety bottom line → efficiency optimization". The core technical feature is as follows: Using CRI < τ (τ is dynamically set according to the safety level, e.g., τ = 0.3 for Level 1, τ = 0.5 for Level 2, and τ = 0.7 for Level 3) as an inviolable hard constraint, all candidate slots are divided into a feasible set S_feasible = {i | CRI(i) < τ} and an infeasible set. An optimization model aimed at minimizing the operational cost (transportation distance D(i) + container handling cost F(i)) is solved only within the feasible set. If the feasible set is empty, an alarm is triggered and the process transitions to a pre-evacuation procedure.

[0058] The core breakthrough of this technology lies in elevating safety from a "soft constraint" to a "hard constraint"—in the ε-constraint method, the safety threshold τ is an absolute admission condition, rather than a weighted factor that can be weighed. Therefore, the system will never allocate a container to a slot where CRI≥τ under any circumstances, thus fundamentally guaranteeing the safety baseline.

[0059] (v) An adaptive pre-evacuation mechanism triggered by environmental thresholds to solve the closed-loop problem of "early warning → action". The core technical feature is that when environmental parameters exceed the level 2 warning threshold (e.g., wind speed ≥ 17.2 m / s, rainfall ≥ 30 mm / h, tide height ≥ 80% of design elevation, seismic acceleration ≥ 0.05g) or the feasible location set is empty, the pre-evacuation process is automatically triggered without manual intervention. Furthermore, the system extrapolates the risk evolution path in a digital twin, identifies a set of high-risk containers V with CRI > τ_evac (τ_evac is dynamically set according to the safety level, e.g., 0.5 for level 1, 0.7 for level 2, and 0.85 for level 3), and solves for a redistribution scheme in the safe area (CRI < τ) with the objective of minimizing the total evacuation cost min Σ(w_i × d_i). It then generates an evacuation instruction sequence according to priority (hazardous goods containers w = 3.0 → heavy containers w = 2.0 → other containers w = 1.0).

[0060] The key innovation of this technology lies in the construction of a complete automated chain of "threshold triggering → risk simulation → high-risk identification → solution solving → instruction issuance", which reduces the pre-evacuation response time from 95 minutes of manual decision-making to 8 minutes, fundamentally solving the defect of existing technologies that "only issue warnings but do not take action".

[0061] In summary, the core technical concept of this invention can be simply summarized as follows: taking "safety is an uncompromising hard constraint" as the fundamental starting point, constructing a closed-loop safety positioning system of "real-time perception → dynamic deduction → coupled evaluation → hard constraint optimization → automatic evacuation", fundamentally solving the systemic defects of existing technologies such as "emphasizing efficiency over safety, having early warnings but no action, and applying a one-size-fits-all approach".

[0062] It should be noted that the breakthrough of this invention lies not in repairing or improving existing technologies (such as adding a safety penalty term to the efficiency optimization function), but in reconstructing the value ranking of location decisions. By elevating safety from "one of the weighable optimization objectives" to "an absolute bottom line that cannot be breached," all other objectives (efficiency, cost) are optimized and designed above this bottom line.

[0063] The following is an illustrative illustration through examples. Example 1: Multimodal Environment Perception and Data Fusion The technical problem to be solved in this embodiment is: how to acquire multi-dimensional environmental parameters of the entire storage yard in real time and accurately, so as to provide a reliable data foundation for subsequent digital twin simulation and risk coupling calculation. Specifically, storage yard environmental parameters have three major challenges: multi-source (various types of parameters such as wind speed, tide level, rainfall, visibility, and seismic acceleration), spatiotemporal heterogeneity (different sampling frequencies and accuracies of different sensors), and noise interference (sensor failure or communication interference may lead to erroneous data).

[0064] The research revealed that while existing meteorological early warning systems (such as CN118863154A) can collect meteorological data, their sensor deployment is relatively sparse (usually only 1-2 meteorological stations are deployed on the roof of the terminal office building or at the lighthouse), failing to reflect the spatial distribution differences of environmental parameters within the yard caused by factors such as container stack obstruction and terrain undulations. Taking wind speed as an example, the actual wind speed on the windward and leeward sides of the yard can differ by 3-5 times. Relying solely on data from a single measuring point will severely overestimate or underestimate the actual wind-induced risks in various areas of the yard.

[0065] Based on this finding, this application proposes deploying a sufficiently dense sensor grid within the storage yard to ensure that the spatial resolution meets the accuracy requirements of digital twin simulation (50-100 meter grid spacing). Simultaneously, since each sensor uses a different time base and sampling frequency, time deviations must be eliminated through NTP timestamp synchronization, measurement noise must be eliminated through Kalman filtering, and finally, an environmental parameter matrix with a unified time resolution will be output.

[0066] This embodiment provides a multimodal environment perception and data fusion scheme, specifically including: (1) Sensor Deployment Scheme: A gridded environmental sensor array is deployed within the stockyard. The sensor layout adopts an N×M regular grid (N is the number of rows, M is the number of columns), and the grid spacing is set to 50-100 meters according to the actual size of the stockyard. The environmental sensors deployed at each grid node include: an ultrasonic anemometer (measuring wind speed and direction at a height of 10 meters, with a sampling frequency of not less than 1Hz, and outputting 10-minute moving average wind speed and instantaneous wind speed); a tide level gauge (deployed on the water-facing side of the stockyard, using radar or pressure level gauges, with a measurement accuracy of ±1cm); a rain gauge (using a tipping bucket rain gauge, with a resolution of 0.1mm, automatically outputting the cumulative rainfall per hour); a visibility sensor (using a transmission or scattering visibility meter, with a measurement range of 10m-10km); and a triaxial acceleration seismic sensor (deployed at key locations on the stockyard foundation, with a sampling frequency of 100Hz or higher). Each sensor uploads data in real time via an industrial Ethernet or LoRa wireless sensor network, with a data acquisition cycle set to 5-10 minutes.

[0067] (2) Data Fusion Scheme: First, all sensor data are timestamped using the NTP network time protocol to eliminate time deviations caused by transmission delays. Then, the 3σ criterion (data deviations from the mean exceeding three times the standard deviation are considered outliers) is used for outlier detection and filtering. Next, the Kalman filter algorithm is used to smooth the environmental parameters. The state vector of the Kalman filter includes wind speed, wind direction, tide height, rainfall, visibility, and seismic acceleration, while the observation vector is the real-time measurement value of each sensor. By recursively updating the state estimate and error covariance matrix, the smoothed environmental parameter vector E(t)={v(t), θ(t), h(t), r(t), d(t), a(t)} is output. This environmental parameter matrix is ​​pushed to the digital twin simulation engine in real time via shared memory or a message queue.

[0068] The technical advantages of this embodiment are as follows: High spatial resolution acquisition of environmental parameters across the entire storage yard is achieved through gridded deployment (grid spacing 50-100 meters); spatiotemporal consistency fusion of multi-source heterogeneous data is achieved through NTP synchronization and Kalman filtering, providing accurate, reliable, and real-time input data for the digital twin simulation engine, ensuring the accuracy of subsequent safe location decisions from the source. The spatial resolution of the sensor array enables the system to capture microclimate differences within the storage yard (such as wind speed differences between the windward and leeward sides), laying a data foundation for the generation of refined risk heat maps.

[0069] Example 2: Construction of Digital Twin 3D Model The technical problem to be solved in this embodiment is: how to accurately characterize the spatial structural factors (drainage capacity, wind protection effect, terrain elevation, and storage status) that affect the environmental risk of the storage yard in a digital twin, so that the digital twin model can not only reflect the current storage status, but also serve as the physical foundation for environmental risk projection.

[0070] Existing digital twin solutions (such as CN119294239A) are mainly used for visualization and mapping of yard conditions and gravitational field simulation. Their models contain relatively limited information dimensions, primarily focusing on factors related to operational efficiency such as container location, stacking layer number, and bay topology. However, environmental risk simulations require more multi-dimensional spatial information: flooding simulations require topographic elevation and drainage network data; wind field simulations require data on wind-resistant structures and the shielding provided by surrounding container stacks; and earthquake simulations require data on foundation bearing capacity.

[0071] Analysis revealed that the drainage network topology and windbreak structure topology of the storage yard have a decisive impact on risk distribution. For example, regarding water accumulation risk, under the same rainfall intensity, the water depth in troughs closer to the drainage inlets is significantly lower than in troughs farther away. Similarly, regarding wind farms, containers on the leeward side of the windbreak net bear only 30%-50% of the wind load on the windward side. Without incorporating these factors into the digital twin model, the risk projection results will be severely distorted.

[0072] refer to Figure 3 This embodiment illustrates a digital twin 3D model construction scheme, which may include four sub-models: (1) Topographic Elevation Model (DEM): Based on LiDAR point cloud data or engineering survey data, a digital elevation matrix of the entire storage yard is generated with a spatial resolution of 0.5m×0.5m, which is used for flooding simulation and bottom elevation query of the storage tank.

[0073] (2) Container stacking status model: The container number, dimensions, weight, hazardous chemical class, current slot coordinates, and stacking layer of each container in the yard are obtained in real time through the yard crane PLC data interface and OCR container number recognition system. The stacking status in the digital twin is updated in real time in a three-dimensional voxel format. The system synchronizes the stacking status data every 30 seconds.

[0074] (3) Drainage network topology model: Based on the design drawings of the yard drainage project, a directed graph topology G=(N, E) of the drainage network is constructed, where N is the set of nodes (including rainwater inlets, inspection wells, drainage pipe sections and outlets), and E is the set of edges. Each edge is labeled with hydraulic parameters such as pipe diameter (mm), slope (%), and Manning roughness coefficient.

[0075] (4) Windproof structure obstruction topology model: The ray tracing algorithm is used to pre-calculate the equivalent windward area reduction factor of each slot in the yard under different wind directions (16 wind directions in 22.5° intervals). The reduction factor ranges from 0 to 1 (0 represents complete obstruction, and 1 represents no obstruction). The model is recalculated and cached at fixed intervals (e.g., early morning every day) based on the latest stacking status.

[0076] The technical advantage of this embodiment lies in the fact that, through the collaborative construction of four sub-models, the digital twin model possesses a complete spatial information foundation to support environmental risk simulation. The DEM provides topographic relief data, the drainage network topology provides runoff calculation capabilities, the windbreak and shielding topology provides wind field attenuation calculation capabilities, and the storage status model provides dynamically updated storage yard layouts. The combination of these four elements enables subsequent CFD wind field simulations, flooding simulations, and seismic response spectrum simulations to be conducted on a physically realistic digital foundation, resulting in more accurate and reliable simulation results.

[0077] Example 3: Generation of Environmental Risk Heat Map The technical problem to be solved in this embodiment is: how to predict the distribution of environmental risks in each spatial unit of the storage yard within a future time window based on the real-time environmental parameter matrix and digital twin model, and generate an environmental risk heat map that can be directly used for site selection decisions.

[0078] Existing digital twin systems (such as CN119294239A) can only simulate the current storage status and operational efficiency indicators, lacking the ability to predict environmental risks. The inventors recognize that storage yard environmental risks exhibit significant spatiotemporal dynamics—wind speed distributions six hours before a typhoon arrives are completely different from those at the time of its arrival; the extent of water accumulation during a one-hour downpour is also completely different from that during a six-hour downpour. Relying solely on current environmental data cannot provide forward-looking guidance for site selection decisions.

[0079] Based on this, this application proposes to simulate wind field, flooding, and seismic risks for future time windows within an acceptable computation time (less than 30 seconds) by running lightweight versions of various physical models. Specifically, wind field simulation is performed using a steady-state CFD solver based on simplified RANS equations, with container stacks as the source of porous media resistance; flooding simulation is performed using a distributed hydrological model based on the SCS-CN runoff curve number method, combined with DEM and drainage network topology; and seismic risk simulation is performed using site response spectrum and container stacking dynamic characteristics.

[0080] refer to Figure 3 This embodiment illustrates an environmental risk simulation scheme, which may include three simulation models: (1) Lightweight CFD wind field model: Based on a simplified steady-state solver of the RANS equations, the model uses container stacks as the source of porous medium resistance and outputs the wind speed distribution field at a height of 10 meters above the horizontal plane of the yard. The solution area covers the entire yard, with a grid size of 2m×2m and a computation time of less than 30 seconds. The model inputs include: reference wind speed v at a height of 10 meters, wind direction θ, and the stacking status of containers in each slot (the equivalent windward area reduction factor is queried through the windproof and shielding topology model).

[0081] (2) SCS-CN-based flooding prediction model: Using real-time rainfall r as input, combined with topographic elevation model and drainage network topology, the model predicts the flooding depth of each tank in the storage yard within a future time window ΔT. The model divides the storage yard surface into several hydrological response units (HRUs), and assigns a CN value to each HRU according to the surface cover type (CN=98 for container storage area, CN=95 for road area, and CN=65 for green belt area). Combining drainage network runoff calculation and the bottom elevation of each tank, the model outputs the predicted flooding depth of each tank. ΔT can be configured as 1 hour, 2 hours, or 6 hours.

[0082] (3) Seismic response spectrum extrapolation model: Based on the site response spectrum and the dynamic characteristics of container stacking, the dynamic amplification factor and overturning risk of containers under different stacking layers are calculated. The model input is the peak ground acceleration a, and the seismic risk value of each slot is calculated by combining the current stacking layer tier of each slot.

[0083] The outputs of the three simulation models above are normalized and then merged into an environmental risk heat map H(x,y,t+ΔT) for each spatial unit of the storage yard, where H takes a continuous value between 0 and 1 (0 represents no risk and 1 represents extremely high risk).

[0084] The technical advantages of this embodiment are as follows: by running three extrapolation models in parallel, it can predict the spatial distribution of environmental risks in the future time window within 30 seconds, generating a risk heat map that can be directly used for site selection decisions. Compared with existing technologies that can only reflect the current state, this embodiment achieves advanced prediction of yard environmental risks, upgrading site selection decisions from "passive response" to "proactive prevention." At the same time, the lightweight solver design keeps the computation time within 30 seconds, meeting the timeliness requirements of real-time terminal operations.

[0085] Example 4: Risk Assessment of Container-Environment Coupling The technical problem to be solved in this embodiment is: how to establish a quantitative coupling relationship between the inherent properties of containers and environmental risk factors, so as to achieve differentiated and refined safety assessment of different types of containers.

[0086] Existing technologies apply uniform safety standards to different types of containers, failing to fully consider the impact of the container's inherent properties on risk. Through analysis of extensive wind tunnel experimental data and historical accident data, the inventors have made the following key findings: (1) The risk of an empty container overturning at the same wind speed is more than 10 times that of a loaded container. Taking a 20-foot container as an example, the weight of an empty container is about 2,200 kg, and the weight of a loaded container is about 30,480 kg. At the same wind speed, the ratio of the wind-induced overturning moment of an empty container to its weight is about 14 times that of a loaded container.

[0087] (2) The risk of secondary disasters after hazardous chemical containers are soaked in water is much higher than that of ordinary containers. Some hazardous chemicals react with water (such as releasing flammable gases or spontaneously combusting upon contact with water), and the risk lies not only in the container itself, but also in the threat to the environment and personnel safety after leakage.

[0088] (3) For each additional stacking layer, the wind-induced overturning moment increases by about 100%. Taking a three-layer stack as an example, the wind load borne by the top box is transferred to the bottom box through the latch. The overturning moment of the bottom box is the sum of the moments of each layer. The higher the number of stacking layers, the more exponentially the risk of overturning increases.

[0089] Based on the above findings, this application constructs a container-environment coupling risk index (CRI) model (reference). Figure 5 (Illustrative diagram) The five dimensions of container quality, equivalent windward area, hazardous chemical risk level, target stacking layer, and bottom elevation are quantitatively coupled with the four dimensions of environmental conditions: wind speed, tide height, rainfall, and seismic acceleration. The weight coefficients of each risk component are determined based on the analytic hierarchy process (AHP).

[0090] This embodiment provides a container-environment coupling risk assessment scheme: (1) Construction of Coupled Risk Feature Vector: When a container enters the yard to be selected for a location (triggered by the gate OCR system or the ship loading plan), the system automatically extracts the attribute data of the container and constructs a coupled risk feature vector R_c={m, S_wind, class_haz, tier, z}. Where m is the container mass, S_wind is the equivalent windward area, class_haz is the hazardous chemical risk level (value 0-9), tier is the target stacking layer number (value 1-6), and z is the bottom elevation of the slot.

[0091] (2) CRI calculation: CRI(i)=α·f_wind(i)+β·f_flood(i)+γ·f_seismic(i)+δ·f_hazmat(i); in: f_wind is the wind-induced overturning risk component, calculated as (equivalent windward area × wind speed squared × stacking height) / (mass × base width × gravitational acceleration), denoted as f_wind = (S_wind × v² × h_tier) / (m × w_base × g). f_flood is the water accumulation risk component, which is calculated as the normalized value of the difference between the predicted water accumulation depth and the bottom elevation of the tank, denoted as f_flood=max(0,d_flood(i)-z) / H_max; f_seismic is the seismic risk component, which is calculated as predicted peak ground acceleration × number of stacked layers / seismic design acceleration, denoted as f_seismic = a×tier / a_max; f_hazmat is the hazardous chemical risk amplification factor, which takes the value of hazardous chemical level / 9, denoted as f_hazmat = class_haz / 9.

[0092] (3) Dynamic switching of weight coefficients: The values ​​of α, β, γ and δ are dynamically switched according to the current warning level, where α, β, γ and δ are weight coefficients and satisfy α+β+γ+δ=1; For example, during a typhoon warning: α=0.5, β=0.2, γ=0.1, δ=0.2; For example, during a rainstorm warning: α=0.1, β=0.6, γ=0.1, δ=0.2; For example, during routine operation: α=0.2, β=0.2, γ=0.1, δ=0.5.

[0093] (4) Spatial matching of the coupled risk feature vector with the environmental risk heat map H(x,y,t+ΔT), query the risk value of the corresponding position in the heat map through the trough coordinate (x,y), substitute it into the above formula to calculate CRI(i), and normalize the result to the 0-1 interval.

[0094] The technical advantage of this embodiment lies in achieving differentiated safety assessments for different types of containers through the CRI model. Taking empty and loaded containers as examples from the background section: under the same wind speed conditions, the f_wind component of an empty container is approximately 14 times that of a loaded container. Therefore, empty containers will be preferentially allocated to the leeward side, low-wind-speed area, while loaded containers can be allocated to more efficient slots while meeting safety thresholds. This effectively solves the problem of a one-size-fits-all approach to safety measures in existing technologies, maximizing operational efficiency while ensuring a safety baseline. Furthermore, the dynamic switching mechanism of weighting coefficients allows the location selection decision to adapt to different environmental risk-dominated scenarios, further enhancing the system's scenario adaptability.

[0095] Example 5: Dynamic Safe Location Selection Decision The technical problem to be solved in this embodiment is: how to optimize operational efficiency within the feasible set under the premise of CRI safety threshold as a hard constraint, so as to achieve a balance between safety and efficiency.

[0096] Existing site selection schemes (such as the D3QN method in CN119047691A and the UCT-MCTS gravity model in CN121211960B) primarily prioritize operational efficiency as the optimization objective, with safety considerations typically treated as soft constraints or post-event remedies. This stems from the fact that traditional optimization theory treats safety and efficiency as two equivalent objectives in a multi-objective optimization problem, compromising them through weighting coefficients. However, this approach has a fundamental problem: safety is an inviolable bottom line—in the event of a safety incident, operations will be completely halted, and efficiency will drop to zero.

[0097] After in-depth analysis, this application proposes that the correct solution is to adopt the ε-constraint method, using the safety index as an inviolable hard constraint (CRI < τ) and the efficiency index as the optimization objective, to find the optimal solution within the safe feasible region. This method can both guarantee the safety baseline and minimize efficiency loss. Specifically, with CRI less than the preset safety threshold τ as the hard constraint, all candidate slots are divided into feasible and infeasible sets, and the operation cost is optimized only within the feasible set.

[0098] refer to Figure 6 This embodiment illustrates a dynamic and secure location selection decision-making scheme: (1) Safety threshold setting: Set the safety threshold τ according to the current safety level of the wharf. Level 1 safety level (during typhoon warning) τ=0.3; Level 2 safety level (during rainstorm warning) τ=0.5; Level 3 safety level (during daily operation) τ=0.7. The safety level will automatically switch according to the weather forecast and the warning information from the maritime department.

[0099] (2) Feasible set partitioning: Taking CRI less than the preset safety threshold τ as an inviolable hard constraint, all candidate slots are divided into a feasible set S_feasible={i | CRI(i)<τ} and an infeasible set.

[0100] (3) Optimization Model Construction and Solution: An optimization model with the objective of minimizing operating costs is constructed only within the feasible set S_feasible: min C = ω_1 × D(i) + ω_2 × F(i). Where D(i) is the horizontal transport distance of the candidate container from the gate or ship berth to slot i, calculated using the shortest path algorithm of the yard road network; F(i) is the cost of container re-turning caused by being assigned to slot i. ω_1 and ω_2 are weight coefficients (default ω_1 = 0.6, ω_2 = 0.4). This optimization model uses the ε-constraint method to solve for the optimal slot i* using an integer programming solver (such as CPLEX or Gurobi) or a heuristic algorithm (such as genetic algorithm or simulated annealing).

[0101] (4) Result distribution: The decision results are distributed to the yard crane PLC and truck dispatching system via MQTT or OPC UA protocol for execution.

[0102] (5) Handling when the feasible set is empty: If the feasible set is empty, an alarm will be automatically triggered and the process will jump to the pre-evacuation process (see Example 6 below for details).

[0103] The technical advantage of this embodiment lies in using the ε-constraint method to treat the safety threshold as a hard constraint, fundamentally guaranteeing the safety baseline for yard operations under extreme weather conditions. Simultaneously, efficiency is optimized only within the safe feasible set, avoiding the possibility of sacrificing safety for efficiency. Simulation experiments show that the safety satisfaction rate of this invention is increased from 45% in the prior art to 100%, while the efficiency loss is only 12-18%, demonstrating the superiority of this embodiment in balancing safety and efficiency.

[0104] Example 6: Adaptive Pre-Evacuation Rescheduling The technical problem to be solved in this embodiment is: how to automatically identify and preventively transfer high-risk containers when environmental parameters exceed the secondary warning threshold or the feasible location set is empty, so as to avoid the defects of slow response and easy omission of manual decision-making.

[0105] Existing early warning systems (such as CN118863154A) can only generate alarms and suspend operations, but cannot automatically translate them into specific evacuation action plans. The inventors recognize that the pre-evacuation window before extreme weather is extremely limited (usually only a few hours), and manually formulating evacuation plans has three inherent drawbacks: first, slow response time (requiring expert consultations, on-site surveys, plan formulation, and approval processes, averaging 95 minutes); second, prone to overlooking high-risk containers (manual inspections cannot cover all slots and accurately calculate the risk value of each container); and third, non-optimal evacuation plans (manually formulated evacuation routes are often not the shortest paths and it is difficult to comprehensively consider the overall optimal allocation of multiple high-risk containers).

[0106] Based on this, the inventors propose that a complete pre-evacuation link must be established, which includes "automatic triggering of environmental thresholds → digital twin risk simulation → automatic identification of high-risk containers → automatic solution of evacuation plan → automatic issuance of instructions", thus automating all manual intervention links.

[0107] refer to Figure 7 The illustration demonstrates the steps of the pre-evacuation process: risk evolution prediction → high-risk container identification → reallocation solution → command issuance. Specifically, this embodiment provides an adaptive pre-evacuation rescheduling scheme: (1) Setting of Level II warning thresholds: Level II warning thresholds include, but are not limited to: 10-minute average wind speed ≥ 17.2 m / s; hourly rainfall ≥ 30 mm; tide height exceeding 80% of the wharf's design elevation; and peak ground acceleration ≥ 0.05 g. The thresholds are determined in accordance with national meteorological standards and wharf seismic design specifications.

[0108] (2) Automatic Triggering and High-Risk Container Identification: The system continuously monitors environmental parameters and compares each parameter in the integrated environmental parameter matrix E(t) in real time. When any environmental parameter exceeds the preset secondary warning threshold, the pre-evacuation process is automatically triggered without manual intervention. After triggering, the system first quickly extrapolates the risk spread path in the digital twin (substituting the current E(t) into the three extrapolation models to predict the evolution of the environmental risk heat map in the next 1-6 hours), and identifies the set of high-risk containers V={v_1, v_2, ..., v_k} in the yard whose CRI exceeds the evacuation threshold τ_evac.

[0109] The evacuation threshold τ_evac is dynamically set according to the current security level: τ_evac=0.5 during Level 1 security, τ_evac=0.7 during Level 2 security, and τ_evac=0.85 during Level 3 security; ensuring that τ_evac is always higher than the security threshold τ of the current security level, to avoid the allocated security slots being repeatedly identified as high-risk slots.

[0110] (3) Solution of the evacuation optimization model: The redistribution scheme is solved in the safe area with the objective of minimizing the total evacuation cost: min Σ_{i=1}^k (w_i×d_i). Where w_i is the evacuation weight of the i-th high-risk container (the weight of hazardous chemical containers is set to 3.0, the weight of heavy containers is set to 2.0, and the weight of other containers is set to 1.0); d_i is the horizontal transportation distance of the container from the original storage location to the newly assigned safe storage location. The constraints of the redistribution scheme include: the newly assigned storage location must be located in the safe area where CRI<τ in the current environmental risk heat map, and the evacuation path must not pass through waterlogged or high-risk areas.

[0111] (4) Priority issuance and execution: After the solution is completed, the system generates an evacuation instruction sequence according to priority order (hazardous chemical containers first, loaded containers second, and empty containers last), and issues it for execution through the yard bridge and truck dispatching system. The evacuation progress and risk evolution trend are displayed in real time on the human-machine interface.

[0112] The technical advantages of this embodiment are as follows: Through an automatic environmental threshold triggering mechanism, the pre-evacuation response time is reduced from an average of 95 minutes for manual decision-making to 8 minutes (-92%); through digital twin risk simulation and CRI calculation, the high-risk container identification rate is increased from 76% to 100%; and through solving the evacuation optimization model, the evacuation completion rate within 4 hours is increased from 68% to 100%. Simulation experiments show that by adopting the pre-evacuation mechanism of this embodiment, cargo damage under extreme weather conditions can be reduced by 60-80%, achieving a qualitative leap from "passive manual response" to "proactive system prevention."

[0113] Example 7: System Overall Architecture and Closed-Loop Operation The technical problem to be solved in this embodiment is: how to organically integrate the aforementioned six major functions of multimodal environment perception, data fusion, digital twin inference, CRI assessment, dynamic location decision-making and adaptive pre-evacuation into a closed-loop system to achieve automated closed-loop operation from environment perception to safe location decision-making.

[0114] Existing technical solutions either only involve single functional modules (such as only weather warning, only digital twin simulation, or only efficiency optimization location selection), lacking systematic integration. The inventors recognized that dynamic safety location selection in a storage yard is a complete chain involving perception, deduction, assessment, decision-making, and execution; the absence of any link will lead to incomplete system functionality. Therefore, it is necessary to design a modular, loosely coupled, and highly cohesive system architecture that enables interconnection between modules through standard data interfaces.

[0115] refer to Figure 1The container terminal yard dynamic safety positioning system based on multimodal environmental perception and digital twin simulation provided by this invention consists of six functional modules. These modules are interconnected through standard data interfaces (such as OPC UA, MQTT, and REST API) to form a closed-loop intelligent safety positioning decision-making system. (1) Multimodal environmental perception layer: Deploy a gridded sensor array, including a gridded ultrasonic anemometer, rain gauge, tide gauge, visibility sensor and triaxial accelerometer seismic sensor, and collect environmental parameters at a cycle of 5-10 minutes through industrial Ethernet or LoRa wireless sensor network.

[0116] Preferably, the environmental sensors deployed at each grid node include, but are not limited to: ultrasonic anemometers (used to measure wind speed and direction at a height of 10 meters, with a sampling frequency of not less than 1Hz, outputting 10-minute moving average wind speed and instantaneous wind speed); tidal level gauges (deployed on the water-facing side of the storage yard, using radar or pressure-type level gauges, with a measurement accuracy of ±1cm); rain gauges (using tipping bucket rain gauges, with a resolution of 0.1mm, automatically outputting cumulative rainfall per hour); visibility sensors (using transmission or scattering visibility meters, with a measurement range of 10m-10km, used for operational risk assessment in foggy weather); and triaxial accelerometer seismic sensors (deployed at key locations in the storage yard foundation, measuring peak ground acceleration and response spectrum, with a sampling frequency of 100Hz or higher).

[0117] (2) Data fusion and preprocessing layer: Receives raw data, performs timestamp synchronization, 3σ criterion outlier filtering and Kalman filtering smoothing, and outputs an environmental parameter matrix with uniform time resolution.

[0118] In implementation, due to differences in sampling frequency and data format among various sensors, the NTP network time protocol is first used to timestamp and standardize all sensor data to eliminate time deviations caused by transmission delays. Then, the 3σ criterion (i.e., data deviating from the mean by more than three times the standard deviation is considered an outlier) is used to detect and filter outliers in the raw data of each sensor, preventing erroneous data caused by sensor malfunctions or communication interference from entering subsequent processing stages. Next, a Kalman filter algorithm is used to smooth various environmental parameters. The state vector of the Kalman filter includes wind speed, wind direction, tide height, rainfall, visibility, and seismic acceleration, while the observation vector is the real-time measurement value of each sensor. By recursively updating the state estimate and error covariance matrix, the smoothed environmental parameter vector E(t) is output. In the preferred example, the time resolution of this environmental parameter matrix is ​​consistent with the sensor acquisition cycle (5-10 minutes) and is pushed to the digital twin simulation engine in real time via shared memory or a message queue.

[0119] (3) Digital twin simulation engine: including terrain elevation model submodule, container stacking status submodule, drainage network topology submodule, wind protection and shielding topology submodule, lightweight CFD wind field simulation submodule, water accumulation and flooding simulation submodule and seismic response spectrum simulation submodule; In the preferred example, the digital twin model includes four sub-models: terrain elevation model (DEM), container stacking status model, drainage network topology model, and windbreak and shielding topology model.

[0120] In the preferred example, after receiving the fused environmental parameter matrix, the digital twin simulation engine runs three simulation models in parallel: a lightweight CFD wind field model, a flood inundation simulation model based on SCS-CN, and a simplified seismic response spectrum simulation model.

[0121] In the preferred example, the outputs of the three major simulation models are normalized and then fused into an environmental risk heat map H(x,y,t+ΔT) for each spatial unit of the storage yard.

[0122] (4) Container-Environment Coupling Risk Assessment Module: Extract attribute data of incoming containers to construct coupling risk feature vectors, and perform spatial matching with environmental risk heat map to calculate CRI index. The weight coefficient of CRI index is dynamically switched according to the current warning level. When a container enters the yard awaiting berth selection (triggered by the gate OCR system or ship loading plan), the system automatically extracts the container's attribute data and constructs a coupled risk feature vector R_c={m, S_wind, class_haz, tier, z}. Subsequently, this feature vector is spatially matched with the environmental risk heatmap H(x,y,t+ΔT) (by querying the risk value at the corresponding location on the heatmap using the slot coordinates (x,y)), and the coupled risk index is calculated according to a weighted multi-factor model.

[0123] (5) Dynamic safety location selection decision module: The feasible set is selected with CRI<τ as a hard constraint. The multi-objective optimization model is solved within the feasible set to output the optimal slot and the selection result is sent to the yard crane PLC and truck scheduling system via MQTT or OPC UA protocol.

[0124] In implementation, a safety threshold τ is first set according to the current terminal safety level; then, with CRI less than the preset safety threshold τ as an inviolable hard constraint, all candidate slots are divided into a feasible set S_feasible={i | CRI(i)<τ} and an infeasible set, and an optimization model aimed at minimizing operating costs is constructed only within the feasible set S_feasible.

[0125] In the preferred example, the optimization model uses the ε-constraint method to solve for the optimal slot i* using an integer programming solver or a heuristic algorithm.

[0126] In addition, the decision results are sent to the yard crane PLC and truck dispatching system for execution via MQTT or OPC UA protocol.

[0127] (6) Adaptive pre-evacuation rescheduling module: When environmental parameters exceed the level 2 warning threshold, it automatically identifies high-risk container sets and solves the redistribution scheme to generate an evacuation instruction sequence.

[0128] During implementation, the system continuously monitors environmental parameters and compares them in real time with each parameter in the fused environmental parameter matrix E(t). When any environmental parameter exceeds the preset secondary warning threshold, the pre-evacuation process is automatically triggered without manual intervention. After triggering, the system first quickly simulates the risk spread path in the digital twin, identifying the set of high-risk containers V={v_1, v_2, ..., v_k} in the yard whose CRI exceeds the evacuation threshold τ_evac. Then, for each high-risk container in V, the system solves for a redistribution scheme in the safe area with the goal of minimizing the total evacuation cost. After the solution is completed, the system generates an evacuation instruction sequence according to priority and issues it for execution through the yard crane and truck scheduling system.

[0129] Figure 1 The system showcases six functional modules: a multimodal environment perception layer, a data fusion and preprocessing layer, a digital twin simulation engine, a container-environment coupling risk assessment module, a dynamic safety location selection decision module, and an adaptive pre-evacuation rescheduling module, as well as the data flow and control flow connections between these modules. Subsequent location selection commands are then sent to the yard crane PLC and the truck dispatching system (not shown in the diagram). These six modules are interconnected through standard data interfaces, forming a closed-loop system from environmental perception to safety location selection decision-making.

[0130] The technical advantages of this embodiment are as follows: the modular design allows each module to be independently upgraded and expanded; the standard data interface ensures the interoperability of the system; and the closed-loop architecture realizes the entire automated process from perception to decision-making to execution. This system architecture has good scalability and can easily integrate new sensor types (such as air quality sensors and radiation sensors) and new inference models (such as fire spread models and hazardous material diffusion models) in the future without requiring a complete refactoring of the overall architecture.

[0131] refer to Figure 2 The method for dynamic safety location selection of container terminal yards based on multimodal environmental perception and digital twin simulation provided by this invention may include the following steps: Step S1: Deploy a gridded multimodal environmental sensor array in the stockyard to collect multidimensional environmental parameters at a preset period, and output a fused environmental parameter matrix after data fusion and preprocessing.

[0132] Preferably, the multimodal environmental sensor array is deployed as an N×M regular grid (N is the number of rows, M is the number of columns), with a grid spacing of 50-100 meters, to ensure that the spatial resolution of environmental parameter acquisition meets the accuracy requirements of digital twin simulation. Multidimensional environmental parameters include: wind speed and direction at a height of 10 meters, tide height, rainfall, atmospheric visibility, and peak ground acceleration (PGA). Data fusion and preprocessing include: timestamping all sensor data using the Network Time Protocol (NTP), outlier detection and filtering using the 3σ criterion, smoothing using a Kalman filter algorithm, and outputting an environmental parameter vector E(t)={v(t), θ(t), h(t), r(t), d(t), a(t)}, where v is wind speed, θ is wind direction, h is instantaneous tide height, r is hourly rainfall, d is atmospheric visibility, and a is PGA.

[0133] Preferably, each sensor uploads data to the data fusion and preprocessing layer in real time via an industrial Ethernet or LoRa wireless sensor network, with the data acquisition cycle set to 5-10 minutes to meet the real-time requirements of location selection decision.

[0134] Preferably, the time resolution of the environmental parameter matrix is ​​consistent with the sensor acquisition cycle (5-10 minutes), and it is pushed to the digital twin inference engine in real time via shared memory or message queue.

[0135] Step S2: Construct and continuously update the digital twin 3D model of the storage yard. The digital twin 3D model includes a terrain elevation model, a real-time container storage status model, a drainage network topology model, and a windproof structure shielding topology model.

[0136] (1) Topographic Elevation Model (DEM): Based on LiDAR point cloud data or engineering survey data, a digital elevation matrix of the entire storage yard is generated with a spatial resolution of 0.5m×0.5m, which is used for flooding simulation and bottom elevation query of the storage tank. (2) Container stacking status model: The container number, size, weight, hazardous chemical class, current slot coordinates and stacking layer of each container in the yard are obtained in real time through the yard crane PLC data interface and OCR container number recognition system, and the stacking status in the digital twin is updated in real time in the form of three-dimensional voxels; (3) Drainage network topology model: Based on the design drawings of the yard drainage project, a directed graph topology of the drainage network is constructed. The nodes include rainwater inlets, inspection wells, drainage pipe sections and outlets. The edges represent the connection relationship of the pipe sections and are labeled with hydraulic parameters such as pipe diameter, slope, Manning roughness coefficient, etc., for the calculation of surface runoff confluence in the simulation of water accumulation and flooding. (4) Windproof structure shading topology model: The ray tracing algorithm is used to pre-calculate the equivalent windward area reduction factor of each slot in the yard under different wind directions due to the shading of surrounding container stacks and buildings. The reduction factor ranges from 0 to 1 (0 represents complete shading, and 1 represents no shading). The model is recalculated and cached at fixed intervals (such as early morning every day) based on the latest stacking status for real-time query during wind field simulation.

[0137] Step S3: Based on the fused environmental parameter matrix, run the environmental risk simulation model in the digital twin engine to predict the spatial distribution of environmental risks within the future time window and generate a heat map of environmental risks in the storage yard.

[0138] Preferably, after receiving the fused environment parameter matrix, the digital twin inference engine runs three inference models in parallel: (a) Lightweight CFD wind field model: Based on a simplified steady-state solver of the RANS equations, the container stacks are used as the source of porous medium resistance terms, and the wind speed distribution field at a height of 10 meters on the horizontal plane of the storage yard is output. The calculation time is less than 30 seconds. (b) SCS-CN-based flooding simulation model: using real-time rainfall as input, combined with topographic elevation model and drainage network topology, the water depth of each tank in the storage yard within ΔT in the future is predicted. ΔT can be configured as 1 hour, 2 hours or 6 hours. (c) Simplified seismic response spectrum extrapolation model: Based on the site response spectrum and container stacking dynamic characteristics, calculate the dynamic amplification factor and overturning risk of containers under different stacking layers.

[0139] The outputs of the three major simulation models are normalized and then merged into an environmental risk heat map H(x,y,t+ΔT) for each spatial unit of the storage yard, where (x,y) are the planar coordinates of the storage yard, and H takes a continuous value between 0 and 1 (0 represents no risk and 1 represents extremely high risk).

[0140] Step S4: When a container arrives at the site to be selected, extract the attribute data of the container and construct a coupled risk feature vector.

[0141] Preferably, when a container enters the terminal awaiting berthing (triggered by the gate OCR system or ship loading plan), the system automatically extracts the container's attribute data and constructs a coupled risk feature vector R_c={m, S_wind, class_haz, tier,z}. Here, m is the container mass (unit: kg, queried from the terminal production system database or default standard value: approximately 2200 kg for an empty 20-foot standard container, and the mass of a loaded container = empty container mass + cargo mass, with a maximum total mass of approximately 30480 kg); S_wind is the equivalent windward area (unit: m²). 2The calculation method is: container frontal projected area × wind pressure center height correction factor. The windward area of ​​a 20-foot container is approximately 7.7m². 2 40 feet is approximately 14.3m 2 When stacking, the windward area of ​​the topmost container is used for calculation); class_haz is the risk level of hazardous chemicals (value 0-9, 0 represents non-hazardous chemicals, 1-9 correspond to Class 9 hazardous chemicals in the International Maritime Dangerous Goods Code IMDG Code, the larger the number, the higher the risk); tier is the target stacking layer number (value 1-6, 1 represents the ground layer, 6 represents the sixth layer); z is the bottom elevation of the tank (unit: m, obtained from the terrain elevation model).

[0142] Step S5: Spatial matching and coupling calculation are performed between the coupled risk feature vector and the environmental risk heatmap to obtain the container-environment coupling risk index CRI(i) for each empty slot in the yard. It should be noted that the CRI calculation process can be referenced from other embodiments and will not be elaborated further.

[0143] Step S6: Using CRI less than a preset safety threshold τ as a hard constraint, divide all candidate slots into a feasible set S_feasible={i | CRI(i)<τ} and an infeasible set; if the feasible set is not empty, solve an optimization model within the feasible set with the objective of minimizing the operation cost, and output the optimal slot; if the feasible set is empty, trigger an alarm and enter the pre-evacuation process. It should be noted that the optimal slot selection process can be referred to the illustrations in other embodiments, and will not be elaborated further.

[0144] Step S7: Compare the fused environmental parameter matrix with the preset secondary warning threshold in real time. When any environmental parameter exceeds the secondary warning threshold, the adaptive pre-evacuation process is automatically triggered to track the risk evolution path in the digital twin and identify the high-risk container set whose CRI exceeds the evacuation threshold τ_evac.

[0145] Preferably, the Level II warning thresholds include, but are not limited to: 10-minute average wind speed ≥ 17.2 m / s (equivalent to Force 8 wind, which has affected the safety of empty container stacking); hourly rainfall ≥ 30 mm (which may cause water accumulation in low-lying areas of the yard); tide height exceeding 80% of the wharf's design elevation (which may cause seawater backflow); and peak ground acceleration ≥ 0.05 g (equivalent to seismic intensity V in China, which may affect the stability of high-stacked containers).

[0146] Preferably, the evacuation threshold τ_evac is dynamically set according to the current safety level: τ_evac=0.5 during Level 1 safety level (typhoon warning), τ_evac=0.7 during Level 2 safety level (rainstorm warning), and τ_evac=0.85 during Level 3 safety level (routine operation); ensuring that τ_evac is always higher than the safety threshold τ of the current safety level, and avoiding the assigned safety slots from being repeatedly identified as high-risk slots.

[0147] Once triggered, the system first rapidly extrapolates the risk contagion path in the digital twin (that is, substitutes the current E(t) into the three extrapolation models to predict the evolution of the environmental risk heat map in the next 1-6 hours), and identifies the set of high-risk containers V={v_1, v_2, ..., v_k} in the yard whose CRI exceeds the evacuation threshold τ_evac.

[0148] Step S8: With the goal of minimizing the total evacuation cost, solve the redistribution scheme for the high-risk container set within the safe area, generate an evacuation instruction sequence, and issue it for execution.

[0149] In practice, for each high-risk container in the set V, the redistribution scheme is solved in the safe area with the goal of minimizing the total evacuation cost: min Σ(w_i×d_i), where w_i is the evacuation weight of the i-th high-risk container (for example, the weight of hazardous chemical containers is set to 3.0, the weight of heavy containers is set to 2.0, and the weight of other containers is set to 1.0, and the weight setting is determined based on the loss of cargo value and the severity of the safety accident); d_i is the horizontal transportation distance (in meters) of the container from the original slot to the newly allocated safe slot.

[0150] Preferably, the evacuation plan also uses integer programming to solve the problem, with additional constraints: the newly assigned slot must be located in the safe zone where CRI < τ in the current environmental risk heat map, and the evacuation route must not pass through waterlogged or high-risk areas.

[0151] After the solution is completed, the system generates an evacuation instruction sequence according to priority (hazardous chemical containers first, loaded containers second, and empty containers last), which is then issued and executed through the yard crane and truck dispatching system. The evacuation progress and risk evolution trend are displayed in real time on the human-machine interface.

[0152] Figure 2 The document demonstrates the sequential execution of eight steps: S1 Environmental Parameter Acquisition → S2 Digital Twin Modeling → S3 Risk Simulation → S4 Container Attribute Extraction → S5 Coupled Risk Calculation → S6 Safe Location Selection Decision → S7 Risk Monitoring and Early Warning → S8 Adaptive Pre-Evacuation.

[0153] The technical solution of the present invention can achieve the following objectives: (1) Multi-dimensional environmental parameters (including wind speed, wind direction, tide height, rainfall, visibility, peak ground acceleration, etc.) of the entire yard are collected in real time by a grid-deployed multi-modal environmental sensor array. After data fusion and preprocessing, a unified environmental parameter matrix is ​​formed, providing a real-time and accurate environmental risk data basis for safe location selection decisions. (2) Construct a high-precision digital twin three-dimensional model of the storage yard, integrate terrain elevation, container storage status, drainage network topology and windproof structure shielding topology, and predict the spatial distribution of environmental risks within the future time window through lightweight CFD wind field simulation, SCS-CN-based water inundation simulation and seismic response spectrum simulation, and generate a storage yard environmental risk heat map. (3) Establish a container-environment coupling risk quantification model. Based on the coupling relationship between the container's own attributes (mass, equivalent windward area, hazardous chemical risk level, target stacking layer, bottom elevation) and real-time environmental conditions, calculate the container-environment coupling risk index (CRI) for each candidate slot to achieve differentiated and refined safety assessment. (4) Construct a multi-objective optimization model with CRI safety threshold as hard constraint and work efficiency as soft objective, and use the ε-constraint method to solve for the optimal safe slot, so as to ensure the safety bottom line while taking into account work efficiency; (5) Establish an adaptive pre-evacuation mechanism triggered by environmental thresholds. When environmental parameters exceed the secondary warning threshold or the feasible set of locations is empty, the risk evolution path is automatically tracked in the digital twin, high-risk container sets are identified, and a redistribution scheme is generated with the goal of minimizing the total evacuation cost, so as to realize the preventive container transfer before the arrival of extreme weather.

[0154] Example 8: A Comprehensive Application Case of Container Arrival Location Selection During Typhoon Weather The following section provides a comprehensive demonstration of the complete application process of various embodiments of the present invention during typhoon weather, using specific scenarios as examples.

[0155] Scenario Description: A coastal container terminal received a blue typhoon warning issued by the meteorological department (it is expected to be affected by the typhoon within the next 48 hours, with an average wind speed of 17.2-20.7 m / s in 10 minutes). The terminal's safety management department set the safety level to Level 1 (τ=0.3) and activated the system of this invention for dynamic safety positioning.

[0156] Technical Problem Description: Before a typhoon arrives, the storage yard faces the threat of strong winds. Different types of containers (hazardous goods containers, loaded containers, and empty containers) have vastly different risks of overturning under the same wind conditions, requiring differentiated safety positioning for each type.

[0157] Research Approach: Through wind tunnel experiments, the inventors discovered that the risk of an empty container overturning at a wind speed of 14 (approximately 25 m / s) is 14 times that of a fully loaded container. If empty and loaded containers are allocated to the same area using the same safety standards, the empty container will face a severe risk of overturning during a typhoon. Therefore, stricter safety constraints must be imposed on the empty container.

[0158] Application process: (1) The multimodal environment perception layer continuously collects environmental parameters according to the scheme described in Example 1. Taking 6 hours before the arrival of the typhoon as an example, the measured 10-minute average wind speed at the yard is v=14.5m / s (level 7 wind), wind direction θ=135° (southeast wind), instantaneous tide height h=2.1m (the design elevation of the wharf is 3.5m, 80% is 2.8m, and the current tide height has not yet triggered a tide disaster warning), hourly rainfall r=0mm (the outer periphery of the typhoon has not yet brought rainfall), atmospheric visibility d=8500m (visibility is good), and peak ground acceleration a=0 (no seismic activity). After the above parameters are smoothed by Kalman filtering in the data fusion and preprocessing layer, the environmental parameter matrix E(t)={14.5, 135°, 2.1, 0, 8500, 0} is output and pushed to the digital twin inference engine.

[0159] (2) The digital twin simulation engine runs the CFD wind field model according to the scheme described in Example 3. Since the current wind direction is 135° (southeast wind) and the typhoon center is located in the southeast direction of the dock, the wind field simulation uses v=14.5m / s and θ=135° as the inlet boundary conditions, and takes the existing container stacks in the yard as the porous medium resistance source term (resistance coefficient C_d=0.8, porosity ε=0.4) to solve the RANS equation to obtain the wind speed distribution field at a height of 10 meters on the horizontal plane of the yard. The simulation results show that the wind speed amplification factor in the southeast area (windward side) of the yard is 1.15-1.25 (meaning the actual wind speed can reach 16.7-18.1 m / s, which is close to level 8 winds). In the northwest area, due to the blocking effect of the stacked containers, the wind speed reduction factor is 0.6-0.75 (the actual wind speed is approximately 8.7-10.9 m / s). There is a windbreak net structure in the northwest corner of the yard, which forms a significant low-speed wake zone on its leeward side (wind speed reduced to less than 0.4 times). The wind field simulation results are output in a 5m×5m grid, with each grid cell recording the predicted wind speed value v(i).

[0160] Meanwhile, the flooding simulation model and the seismic response spectrum simulation model were running simultaneously. Since the current hourly rainfall r=0, the flooding simulation results show that the water depth at each trough is 0; since there is no seismic activity, the seismic simulation results show that f_seismic(i)=0 at each trough. Therefore, the current environmental risk heatmap H(x,y,t+ΔT) is mainly determined by the wind field risk component.

[0161] (3) Assume three container trucks carrying different types of containers enter the yard through the gate, and yard spaces need to be allocated to them. The container information of the three trucks is as follows: Container A is a 20-foot hazardous materials container (IMDG Class 3, flammable liquid, class_haz=3), with a mass m=24000kg and an equivalent windward area S_wind=7.7m. 2 The container is a 40-foot standard heavy container (non-hazardous materials, class_haz=0), with a mass m=28000kg and an equivalent windward area S_wind=14.3m². 2 Container C is a 20-foot empty container (non-hazardous materials, class_haz=0), with a mass m=2200kg and an equivalent windward area S_wind=7.7m². 2 The system extracts the aforementioned attribute data according to step S4 and constructs coupling risk feature vectors for each bin: R_c,A={24000, 7.7, 3, tier,z}, R_c,B={28000, 14.3, 0, tier, z}, and R_c,C={2200, 7.7, 0, tier, z}.

[0162] In other words, the information of the three truck trailers carrying different types of containers is summarized as follows: Container A (20-foot hazardous materials container, class_haz=3, m=24000kg, S_wind=7.7m²); Container B (40-foot ordinary loaded container, class_haz=0, m=28000kg, S_wind=14.3m²); Container C (20-foot empty container, class_haz=0, m=2200kg, S_wind=7.7m²).

[0163] According to step S5, the system queries all currently available slots (assuming a total of 156 available slots) and calculates the CRI value of the three containers mentioned above for each available slot. Taking three representative slots as examples: Slot P1 is located in the southeast windward area of ​​the yard (coordinates correspond to the wind speed calculated from the wind field, v=17.8m / s), bottom elevation z=1.2m, maximum allowed stacking layer tier_max=5, currently available (tier=1); Slot P2 is located in the leeward area of ​​the windbreak net on the northwest side of the yard (v=5.8m / s), bottom elevation z=1.2m, tier_max=5; Slot P3 is located in the middle of the yard (v=12.3m / s), bottom elevation z=1.1m (slightly lower than P1 and P2), tier_max=5. When calculating CRI, we take α=0.5, β=0.2, γ=0.1, δ=0.2 (Level 1 safety), f_flood=0, f_seismic=0.

[0164] Then, calculate the CRI value according to the scheme described in Example 4: For container A (a hazardous materials container), its CRI is 0.209 in slot P1 (windward area, v=17.8m / s), 0.082 in slot P2 (leeward area, v=5.8m / s), and 0.135 in slot P3 (central area, v=12.3m / s). For container B (a regular loaded container), its CRI is 0.154 at P1, 0.016 at P2, and 0.074 at P3. For container C (empty container), its CRI is 1.55 in P1, 0.164 in P2, and 0.74 in P3.

[0165] The specific calculations are as follows: For container A (hazardous materials container): CRI(A,P1)=0.5×(7.7×17.8 2 ×2.59) / (24000×2.35×9.8)+0.2×0+0.1×0+0.2×(3 / 9)=0.5×0.284+0+0+0.067=0.209; CRI(A,P2)=0.5×(7.7×5.8 2 ×2.59) / (24000×2.35×9.8)+0.067=0.5×0.030+0.067=0.082; CRI(A,P3)=0.5×(7.7×12.3 2 ×2.59) / (24000×2.35×9.8)+0.067=0.5×0.136+0.067=0.135; For container B (a standard loaded container): CRI(B,P1)=0.5×(14.3×17.8 2 ×2.59) / (28000×2.35×9.8)+0+0+0=0.5×0.307=0.154; CRI(B,P2)=0.5×(14.3×5.8 2 ×2.59) / (28000×2.35×9.8)=0.5×0.033=0.016; CRI(B,P3)=0.5×(14.3×12.3 2 ×2.59) / (28000×2.35×9.8)=0.5×0.147=0.074; For container C (empty container): CRI(C,P1)=0.5×(7.7×17.8 2 ×2.59) / (2200×2.35×9.8)=0.5×3.10=1.55; CRI(C,P2)=0.5×(7.7×5.8 2 ×2.59) / (2200×2.35×9.8)=0.5×0.329=0.164; CRI(C,P3)=0.5×(7.7×12.3 2 ×2.59) / (2200×2.35×9.8)=0.5×1.48=0.74.

[0166] (4) The feasible set was partitioned according to the scheme described in Example 5 with τ=0.3 as a hard constraint. The results showed that the feasible set of container A included P1 (CRI=0.209<0.3), P2 (0.082), and P3 (0.135); the feasible set of container B also included P1, P2, and P3; the feasible set of container C only included P2 (CRI=0.164<0.3), and the CRIs of P1 and P3 both exceeded 0.3 (1.55 and 0.74 respectively), so they were excluded. Subsequently, the system solved the optimization model with the goal of minimizing the overall operating cost within the feasible set. Assuming the horizontal transport distances from the gate to P1, P2, and P3 are D(P1) = 350m, D(P2) = 680m, and D(P3) = 420m respectively, and the turnover cost F is 0 for all slots (assuming there are no already stacked containers below each slot or the stacking time is matched), then the optimal slot for container A is P1 (C = 0.6 × 350 = 210), and the optimal slot for container B is also P1. However, since P1 can only be allocated to one container, the system adopts a greedy allocation strategy: it prioritizes allocating containers with lower CRI (greater safety margin). Container A has a CRI(P1) of 0.209, and container B has a CRI(P1) of 0.154. Container B has a larger safety margin (0.3 - 0.154 = 0.146 > 0.3 - 0.209 = 0.091). Therefore, P1 is allocated to container B, and container A is allocated to the second-best P3 (C = 0.6 × 420 = 252). Container C can only be assigned to P2 (C=0.6×680=408). The final allocation result is: Container B→P1 (southeast windward area, but the loaded container has strong wind resistance), Container A→P3 (central area, hazardous chemical containers are allocated to medium-risk areas), Container C→P2 (northwest leeward area, empty containers have weak wind resistance, allocated to the safest area). This result fully demonstrates the differentiated safety allocation capability of this invention: although the hazardous chemical container is allocated to the medium-risk area, its CRI=0.135 is far less than the threshold of 0.3, with sufficient safety margin; the empty container, due to its light weight and weak wind resistance, is strictly restricted to the low-risk area (slot P2), avoiding the risk of overturning in strong winds.

[0167] In summary, the feasible sets for containers A and B include P1, P2, and P3; the feasible set for container C only includes P2 (the CRIs of both P1 and P3 exceed 0.3). Considering the transportation distance and the cost of repositioning, the final placement results are: Container B → P1 (loaded containers have strong wind resistance), Container A → P3 (hazardous materials containers are allocated to medium-risk areas with sufficient safety margin), Container C → P2 (empty containers have weak wind resistance and are forcibly allocated to the safest leeward area).

[0168] This embodiment fully demonstrates the differentiated safety placement capability of the present invention: empty containers, due to their light weight and weak wind resistance, are strictly confined to the low-risk area (slot P2) to avoid the risk of overturning in strong winds; although hazardous chemical containers are assigned to the medium-risk area, their CRI=0.135 is far less than the threshold of 0.3, providing ample safety margin; loaded containers, due to their strong wind resistance, are assigned to the windward area to shorten transportation distances and improve operational efficiency. Therefore, the safety satisfaction rate reaches 100%, with an operational efficiency loss of only about 14%.

[0169] Example 9: A Comprehensive Application Case of Adaptive Pre-evacuation in Rainstorm and Flooding Scenarios The following section, in conjunction with specific scenarios, comprehensively demonstrates the complete application process of adaptive pre-evacuation in the context of heavy rain and flooding.

[0170] Scenario Description: This scenario describes the complete process of preventatively transferring high-risk containers in low-lying areas of a container terminal during a severe rainstorm, utilizing the adaptive pre-evacuation function of the method described in this invention. A container terminal is hit by a severe rainstorm, with an hourly rainfall of r=45mm (exceeding the level 2 warning threshold of 30mm), and the rainfall is expected to continue for more than 4 hours, posing a serious risk of flooding to low-lying areas of the container terminal. The safety level is level 2 (τ=0.5), and the evacuation threshold τ_evac=0.7.

[0171] Technical Issue Description: The low-lying area in the southeast corner of the yard has an elevation of only 0.8-1.0m, making it the lowest point in the entire yard. Under continuous heavy rainfall, this area will accumulate water to a depth of 0.45-0.55m, and some containers with a bottom elevation below 1.0m will be at risk of being submerged. High-risk containers in this area must be moved to a safe area before they are submerged.

[0172] Research Approach: Existing technologies (such as CN118863154A) can only generate rainstorm warnings and suspend operations based on meteorological data, but cannot automatically generate evacuation plans. If manual decision-making is used, dispatchers need to determine, one by one, which containers need to be evacuated, to which storage locations, and in what order, based on floodwater simulation results. In emergency situations, manual decision-making takes an average of 95 minutes, is prone to overlooking high-risk containers, and may not result in optimal evacuation routes.

[0173] The adaptive pre-evacuation mechanism proposed in this application solves this problem: when the hourly rainfall r=45mm exceeds the secondary warning threshold (30mm), the system automatically triggers the pre-evacuation process; the flood inundation simulation model predicts the water depth distribution 4 hours later, the system automatically identifies the set of high-risk containers with CRI>τ_evac, establishes an assignment problem optimization model to solve the optimal redistribution scheme, and generates an instruction sequence according to the priority of hazardous chemical containers.

[0174] Application process: (1) The multimodal environment perception layer continuously collects environmental parameters according to step S1. At this time, the measured hourly rainfall r=45mm, the 10-minute average wind speed v=5.2m / s (light breeze, not affecting wind field risk), the wind direction θ=90° (east wind), the instantaneous tide height h=2.6m (tide disaster warning not yet triggered), the atmospheric visibility d=1200m (dense fog, but not affecting water accumulation risk), and the peak ground acceleration a=0. After the above parameters are processed by the data fusion and preprocessing layer, the environmental parameter matrix E(t)={5.2, 90°,2.6, 45, 1200, 0} is output and pushed to the digital twin inference engine.

[0175] During implementation, the multimodal environmental perception layer collected an hourly rainfall of r=45mm, which exceeded the level 2 warning threshold (30mm), and the system automatically triggered the pre-evacuation process.

[0176] (2) The digital twin simulation engine runs the SCS-CN flooding simulation model according to the scheme described in Example 3. The model divides the surface of the storage area into more than 5,000 5m×5m hydrological response units (HRUs), and assigns a CN value to each HRU according to the surface cover type: CN=98 for container storage area (surface is reinforced concrete pavement with extremely low permeability); CN=95 for road area; CN=65 for green belt area. The model uses r=45mm / h as input, and combines the terrain elevation model (DEM resolution 0.5m×0.5m) and drainage network topology model to calculate the surface runoff generation and water depth. The simulation results show that the predicted water accumulation depth (d_flood) in the southeast corner of the storage yard (elevation z=0.8-1.0m, the lowest depression area in the entire yard) is 0.25-0.35m after 1 hour; the predicted water accumulation depth (d_flood) in the central area of ​​the storage yard (elevation z=1.2-1.4m) is 0.05-0.15m; and there is no predicted water accumulation in the northwest area of ​​the storage yard (elevation z=1.6-2.0m). Meanwhile, wind field and seismic simulation results show that the wind-induced overturning risk and seismic risk in each area are both zero (low wind speed, no earthquake). Therefore, the current environmental risk heat map H(x,y,t+ΔT) is mainly determined by the water accumulation risk component.

[0177] (3) The system continuously monitors environmental parameters and finds that the hourly rainfall r=45mm has exceeded the level 2 warning threshold of 30mm, automatically triggering the pre-evacuation process. The digital twin simulation engine further simulates the evolution of water accumulation risk in the next 4 hours: Assuming the rainfall continues, the water depth in the low-lying area in the southeast corner of the yard will reach 0.45-0.55m after 4 hours, which exceeds the maximum designed water depth H_max=0.5m. Some containers (especially those with a bottom elevation of less than 1.0m) will face the risk of being soaked. Based on the predicted heat map sequence, the system identifies all containers in the yard with CRI>τ_evac (at this time, it is level 2 safety, τ_evac=0.7) and forms a high-risk container set V.

[0178] Calculations showed that 12 containers located in the low-lying area in the southeast corner of the yard (bottom elevation z < 1.0m, predicted water depth > 0.45m after 4 hours, CRI = 0.6-0.85) were included in V. These included: 2 hazardous chemical containers (Class 8, corrosive substances, class_haz = 8), 5 loaded containers (weighing 20,000-30,000 kg), and 5 empty containers. The system prioritized V according to evacuation priority: the 2 hazardous chemical containers had the highest priority (weight w = 3.0), followed by the 5 loaded containers (w = 2.0), and the 5 empty containers had the lowest priority (w = 1.0). A list of high-risk containers was generated and pushed to the dispatch center's human-machine interface.

[0179] Therefore, the system identified 12 high-risk containers (2 chemical containers, 5 loaded containers, and 5 empty containers) with CRI>0.7 (τ_evac=0.7) in the yard, forming a high-risk set V.

[0180] (4) The system aims to minimize the total evacuation cost and solves the redistribution scheme for the 12 high-risk containers in V. First, the system determines the current safe zone based on the latest environmental risk heat map: it queries all slots where CRI < τ (the safety level has been upgraded to level 2, τ = 0.5) and the bottom elevation z > 1.5m (to avoid the risk of secondary water accumulation). A total of 38 safe and vacant slots are found, mainly distributed in the northwest side and the central higher area of ​​the yard. Then, the assignment problem optimization model is constructed: min Z = Σw_i × d_i, where d_i is the horizontal transportation distance of the i-th high-risk container from the original slot to the new slot. Taking a hazardous chemical container as an example: the original container location coordinates are (120, 45) (low-lying area in the southeast corner), and the nearest safe and available container location coordinates are (35, 180) (higher area in the northwest), with a transportation distance d≈420 meters; the original container location for a loaded container is (115, 50), and the nearest safe container location is (40, 160), with a distance d≈380 meters; the original container location for an empty container is (125, 42), and the nearest safe container location is (50, 150), with a distance d≈350 meters. The Hungarian algorithm was used to solve the assignment problem. The solution time was about 0.8 seconds (12 boxes × 38 candidate slots = 456 variables, which is relatively small). The optimal allocation scheme was output as follows: 2 hazardous chemical boxes were allocated to slots (38, 185) and (42, 190) in the northwest safe area, 5 heavy boxes were allocated to slots (75, 165)-(95, 170) in the central high area, and 5 empty boxes were allocated to the remaining safe slots in the northwest.

[0181] The total evacuation cost Z = 3.0 × 420 + 3.0 × 435 + 2.0 × (380 + 365 + 390 + 375 + 370) + 1.0 × (350 + 340 + 355 + 345 + 360) = 1230 + 4870 + 1750 = 7850 (cost unit: weight-meter).

[0182] The above calculations can be performed using the scheme described in Example 6 to solve the redistribution scheme: Determine the safe zones (slots with CRI < 0.5 and elevation z > 1.5m, totaling 38), construct an optimization model min Σ(w_i × d_i), and solve it using the Hungarian algorithm. Hazardous material containers are assigned to the northwest safe zone, heavy containers to the centrally elevated zone, and empty containers to the remaining safe slots. The total evacuation cost Z = 7850 weight-meters.

[0183] (5) The system generates an evacuation instruction sequence according to priority: first, evacuate 2 hazardous chemical containers (expected to be completed within 30 minutes), then evacuate 5 heavy containers (expected to be completed within 60 minutes), and finally evacuate 5 empty containers (expected to be completed within 45 minutes). The total estimated evacuation time is about 2 hours and 15 minutes, and all evacuations can be completed within the window of 4 hours of continuous rain.

[0184] The evacuation order was sent to 3 yard cranes and 8 container trucks via the MQTT protocol for execution, and the dispatch center's human-machine interface displayed the evacuation progress of each container in real time.

[0185] Ultimately, before the peak of the rainstorm arrived, all 12 high-risk containers were safely transferred to higher ground, preventing the cargo from being soaked and the containers from being damaged. It was estimated that the potential cargo loss was reduced by approximately 3.2 million yuan.

[0186] This embodiment demonstrates the complete application of the adaptive pre-evacuation mechanism in a rainstorm scenario: the entire pre-evacuation process, from triggering to generating an evacuation plan, takes only 8 minutes (compared to 95 minutes for manual decision-making), achieving a 100% high-risk container identification rate (compared to 76% for manual decision-making), a 100% evacuation completion rate within 4 hours (compared to 68% for manual decision-making), and ultimately achieving zero cargo loss (compared to 850,000 yuan in cargo loss under manual decision-making). Therefore, the adaptive pre-evacuation mechanism can reduce the pre-evacuation response time from hours to minutes, significantly reducing the risk of cargo loss and safety accidents under extreme weather conditions.

[0187] Example 10: A Comprehensive Application Example of Daily Location Selection Based on Coupling of Multiple Environmental Factors This scenario describes a container terminal using the method described in this invention to perform multi-environmental factor coupling for location selection under normal weather conditions (light wind, low tide, no rainfall), demonstrating how the algorithm achieves a balance between safety constraints and operational efficiency. Assume a container terminal is in normal operation with clear weather. Measured environmental parameters are: 10-minute average wind speed v = 3.2 m / s (Level 2 wind, wind-induced capsizing risk can be ignored), wind direction θ = 270° (westerly), instantaneous tide height h = 1.2 m (low tide), hourly rainfall r = 0 mm, atmospheric visibility d = 15000 m, and peak ground acceleration a = 0. The safety level is Level 3 (τ = 0.7).

[0188] First, the digital twin simulation engine runs three simulation models. The wind field simulation results show that the wind speed v(i) in each region is between 3.2 and 4.5 m / s (the risk of wind-induced overturning is negligible); the water accumulation simulation results show that d_flood(i) = 0 for each tank location; and the earthquake simulation results show that f_seismic(i) = 0. Therefore, the baseline CRI values ​​of each grid cell in the environmental risk heatmap H(x,y,t+ΔT) are close to 0, indicating relatively lenient safety constraints.

[0189] Subsequently, assuming a batch of 8 containers needs to pass through the gate for storage, their attribute data is as follows: Container 1: 20-foot standard heavy container, m=26000kg, S_wind=7.7m 2 class_haz=0; Container 2: 40-foot standard heavy container, m = 30000 kg, S_wind = 14.3 m 2class_haz=0; Container 3: 20-foot hazardous materials container (Class 5.1, oxidizing substances), m = 22000 kg, S_wind = 7.7 m 2 class_haz=5; Box 4: 40-foot empty box, m = 3800 kg, S_wind = 14.3 m 2 class_haz=0; Box 5: 20-foot standard heavy box, m=24000kg, S_wind=7.7m 2 class_haz=0; Container 6: 20-foot hazardous materials container (Class 2.1, flammable gas), m=21000kg, S_wind=7.7m 2 , class_haz=2; Box 7: 40-foot standard heavy box, m=29000kg, S_wind=14.3m 2 class_haz=0; Box 8: 20-foot empty box, m = 2200 kg, S_wind = 7.7 m 2 , class_haz=0.

[0190] The system extracts the aforementioned attribute data according to step S4 and constructs the coupling risk feature vector for each box.

[0191] According to step S5, the system calculates the CRI value of each container in all available slots (assuming a total of 215). Taking α=0.2, β=0.2, γ=0.1, and δ=0.5 (Level 3 safety), since the current environmental risk is low, the CRI value of each container is much less than the threshold τ=0.7, and the feasible set S_feasible contains all 215 available slots. Taking container 3 (hazardous materials container, class_haz=5) as an example, its CRI calculation result is: CRI(container 3, any slot) = 0.2 × f_wind + 0.2 × 0 + 0.1 × 0 + 0.5 × (5 / 9) = 0.2 × minimum value + 0.278 ≈ 0.278 (f_wind is extremely small due to low wind speed); this value is much less than 0.7, therefore all slots are feasible slots. Similarly, the CRI for box 6 (class_haz=2) is approximately 0.2 × minimum value + 0.5 × (2 / 9) = 0.111, which is even smaller. This indicates that during daily operation, safety constraints are relatively lenient, and location decisions can consider operational efficiency more.

[0192] Following step S6, the system solves an optimization model for all 215 available slots, aiming to minimize the overall operating cost. The overall operating cost function is C(i) = ω_1 × D(i) + ω_2 × F(i) + ω_3 × S(i), where ω_1 = 0.5, ω_2 = 0.3, and ω_3 = 0.2 (efficiency weights are higher during daily operation). The following uses containers 1 and 2 as examples to illustrate the solution process: Container 1 is a 20-foot container, its destination port is Shanghai, and the estimated storage time is 4 days; Container 2 is a 40-foot container, its destination port is Ningbo, and the estimated storage time is 7 days. The system queries the TOS database for the distance D(i) from each available slot to the gate, the estimated container retrieval time of existing containers below the slot (used to calculate the container handling cost F(i)), and the current storage density of the container area to which the slot belongs (used to calculate the container area balance penalty S(i)). Assuming that calculations show the three candidate storage locations with the lowest overall costs are: location Q1 (D=280m, F=0, S=0.1, C=0.5×280+0.3×0+0.2×0.1=140.02); location Q2 (D=320m, F=0, S=0.05, C=160.01); and location Q3 (D=250m, F=1, S=0.2, C=125+30+4=159). Although Q3 has the shortest transport distance, its overall cost is higher than Q1 because of its container handling cost F=1 (an existing container below is expected to be picked up earlier than container 1). Therefore, the system selects Q1 as the optimal storage location for container 1. For container 2 (40-foot container), the system further filters for storage locations that only allow 40-foot containers to be stacked (some storage locations can only stack 20-foot containers due to length limitations), and solves for the optimal solution among the qualified storage locations. Ultimately, the optimal slot allocation for the eight containers was issued and executed through the TOS system. Actual operation verified that the average transport distance for this batch of containers was 310 meters, the turnover rate was 0.8% (only one of the eight containers had a turnover risk, which was avoided through optimized allocation), and the difference in container density was less than 5%, achieving a good balance between safety and efficiency. Compared to the comparative example (using the purely efficiency-optimized D3QN slot selection method, without considering safety constraints), the slot selection results of this invention also demonstrate excellent safety assurance: the CRI values ​​of all hazardous chemical containers were less than 0.3 (far exceeding the safety threshold requirement of 0.7), and in the event of sudden extreme weather events such as typhoons, the slots where these hazardous chemical containers were located already had a high safety margin, eliminating the need for emergency evacuation. This further verifies the effectiveness of the invention's pre-set safety margin technology during daily operation.

[0193] Example 11: Comparative Experiment and Economic Benefit Analysis To verify the technical advantages of this invention over existing technologies, quantitative data support is needed through comparative experiments.

[0194] The Shanghai International Port Group's D3QN method (CN119047691A, representative efficiency optimization location scheme) and the manual decision-making pre-evacuation scheme were selected as controls. Under the same experimental conditions, their safety performance, operational efficiency, and evacuation effect were tested respectively.

[0195] Comparative Experiment 1: This invention vs. Shanghai International Port Group's D3QN method (CN119047691A) Experimental conditions: A coastal container terminal, under a blue typhoon warning (wind speed 17.2-20.7 m / s). Experimental subjects: 100 incoming containers (including 10 hazardous chemical containers, 20 empty containers, and 70 ordinary loaded containers).

[0196] The experimental results are shown in the table below:

[0197] Experimental conclusion: The present invention is significantly superior to the existing technology in terms of safety performance (the safety satisfaction rate is increased from 45% to 100%, and the number of empty container tipping risk exceedances is reduced from 8 to 0), and the efficiency loss is controllable (the average transport distance is increased by 14%).

[0198] Comparative Experiment 2: This Invention vs. Manual Decision-Making Pre-Evacuation Experimental conditions: A red rainstorm warning was issued for a certain port (hourly rainfall of 45 mm). Experimental subjects: 500 containers in the yard.

[0199]

[0200] Experimental conclusion: The pre-evacuation mechanism of this invention is significantly superior to manual decision-making in terms of response speed (reduced from 95 minutes to 8 minutes), recognition accuracy (increased from 76% to 100%), and evacuation completion rate (increased from 68% to 100%).

[0201] Comparative Experiment 3: Economic Benefit Analysis Based on annual operational data from a coastal container terminal:

[0202] The comparative experimental data above show that the present invention is significantly superior to existing technologies in key indicators such as safety satisfaction rate, pre-evacuation response time, high-risk container identification rate, evacuation completion rate, and cargo damage control. The safety satisfaction rate is increased from 45% to 100% (+55 percentage points), the pre-evacuation response time is shortened from 95 minutes to 8 minutes (-92%), and cargo damage in extreme weather is reduced by 60-80%. Moreover, the present invention has significant economic benefits, with an investment payback period of 2.5 years and a 5-year ROI of 180%, demonstrating good commercial feasibility.

[0203] While specific embodiments of this application have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this application is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this application, but all such changes and modifications fall within the scope of protection of this application.

[0204] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic safety location selection of container terminal yards based on multimodal environmental perception and digital twin inference, characterized in that, Includes the following steps: Step S1: Deploy a gridded multimodal environmental sensor array in the stockyard to collect multidimensional environmental parameters at a preset period, and output a fused environmental parameter matrix after data fusion and preprocessing; Step S2: Build and continuously update the digital twin 3D model of the stockpile; Step S3: Based on the fused environmental parameter matrix, run the environmental risk simulation model in the digital twin engine to predict the spatial distribution of environmental risks within the future time window and generate a heat map of environmental risks in the storage yard. Step S4: When a container arrives at the site to be selected, extract the attribute data of the container and construct a coupled risk feature vector; Step S5: Spatial matching and coupling calculation are performed between the coupled risk feature vector and the environmental risk heat map to obtain the container-environment coupling risk index CRI(i) for each empty slot in the yard; Step S6: Using CRI less than a preset safety threshold τ as a hard constraint, divide all candidate slots into a feasible set S_feasible={i | CRI(i)<τ} and an infeasible set; if the feasible set is not empty, solve the optimization model with the goal of minimizing the operation cost within the feasible set and output the optimal slot; if the feasible set is empty, trigger an alarm and enter the pre-evacuation process.

2. The method according to claim 1, characterized in that, The method further includes: Step S7: Compare the fused environmental parameter matrix with the preset secondary warning threshold in real time. When any environmental parameter exceeds the secondary warning threshold, the adaptive pre-evacuation process is automatically triggered to track the risk evolution path in the digital twin and identify the high-risk container set whose CRI exceeds the evacuation threshold τ_evac. Step S8: With the goal of minimizing the total evacuation cost, solve the redistribution scheme for the high-risk container set within the safe area, generate an evacuation instruction sequence, and issue it for execution.

3. The method according to claim 2, characterized in that, In step S7, the secondary warning threshold includes: 10-minute average wind speed ≥ 17.2 m / s; hourly rainfall ≥ 30 mm; tide height exceeding 80% of the wharf's design elevation; peak ground acceleration ≥ 0.05 g; and / or, the evacuation threshold τ_evac is dynamically set according to the current safety level: evacuation threshold τ_evac = 0.5 for the first safety level, 0.7 for the second safety level, and 0.85 for the third safety level, and τ_evac is always higher than the safety threshold τ of the current safety level; And / or, in step S8, the objective function of the total evacuation cost is min Σ (w_i×d_i), where w_i is the evacuation weight of the i-th high-risk container, and d_i is the horizontal transport distance of the container from its original storage location to the newly assigned safe storage location; and / or, the constraints of the redistribution scheme include: the newly assigned storage location is located within the safe area of ​​CRI<τ in the current environmental risk heat map, and the evacuation path does not pass through waterlogged areas or high-risk areas; and / or, the evacuation instruction sequence is generated in priority order, with the priority order being: hazardous chemical containers first, loaded containers second, and empty containers last.

4. The method according to claim 1, characterized in that, The multimodal environmental sensor array described in step S1 is deployed as an N×M regular grid with a grid spacing of 50-100 meters; and / or, the multidimensional environmental parameters include: wind speed and direction at a height of 10 meters, tide height, rainfall, atmospheric visibility, and peak ground acceleration; and / or, the data fusion and preprocessing include: timestamping all sensor data using the Network Time Protocol, using the 3σ criterion for outlier detection and filtering, using the Kalman filter algorithm for smoothing, and outputting an environmental parameter vector E(t)={v(t), θ(t), h(t), r(t), d(t), a(t)}, where v is wind speed, θ is wind direction, h is instantaneous tide height, r is hourly rainfall, d is atmospheric visibility, and a is peak ground acceleration.

5. The method according to claim 1, characterized in that, In step S2, the digital twin 3D model includes a terrain elevation model, a real-time container stacking status model, a drainage network topology model, and a windproof structure shielding topology model. The terrain elevation model generates a digital elevation matrix of the entire storage yard based on LiDAR point cloud data or engineering survey data, with a spatial resolution of 0.5m × 0.5m. The real-time container stacking status model obtains the container number, size, weight, hazardous chemical class, current slot coordinates and stacking layer of each container in the yard in real time through the yard crane PLC data interface and OCR container number recognition system, and updates the stacking status in the digital twin in real time in a three-dimensional voxel manner. The drainage network topology model is a directed graph topology of the drainage network constructed based on the design drawings of the yard drainage project. The nodes include rainwater inlets, inspection wells, drainage pipe sections and outlets, and the edges are labeled with pipe diameter, slope and Manning roughness coefficient. The windproof structure shading topology model uses a ray tracing algorithm to pre-calculate the equivalent windward area reduction factor for each storage location in the yard under different wind directions due to the shading of surrounding container stacks and buildings.

6. The method according to claim 1, characterized in that, The environmental risk simulation model mentioned in step S3 includes: A lightweight CFD wind field model, based on a simplified steady-state solver of the RANS equations, uses container stacks as the source of porous medium resistance terms, and outputs the wind speed distribution field at a height of 10 meters above the horizontal plane of the storage yard. The flooding simulation model based on SCS-CN uses real-time rainfall as input and combines topographic elevation model and drainage network topology to predict the water depth of each tank in the storage yard within the future time window ΔT. The seismic response spectrum extrapolation model, based on the site response spectrum and the dynamic characteristics of container stacking, calculates the dynamic amplification factor and overturning risk of containers under different stacking layers; The environmental risk heat map H(x,y,t+ΔT) is generated by fusing the outputs of the above three extrapolation models after normalization, where (x,y) are the planar coordinates of the stockpile, and H takes a continuous value between 0 and 1.

7. The method according to claim 1, characterized in that, In step S4, the coupled risk feature vector R_c={m,S_wind, class_haz, tier, z} is used, where m is the container mass, S_wind is the equivalent windward area, class_haz is the hazardous chemical risk level, tier is the target stacking layer number, and z is the bottom elevation of the tank. And / or, the formula for calculating the container-environment coupling risk index CRI(i) in step S5 is: CRI(i)=α·f_wind(i)+β·f_flood(i)+γ·f_seismic(i)+δ·f_hazmat(i) Wherein, f_wind is the wind-induced overturning risk component, calculated as (equivalent windward area × wind speed squared × stacking height) / (mass × base width × gravitational acceleration); f_flood is the water accumulation risk component, calculated as the normalized value of the difference between the predicted water accumulation depth and the bottom elevation of the tank; f_seismic is the seismic risk component, calculated as the predicted peak ground acceleration × number of stacking layers / seismic design acceleration; f_hazmat is the hazardous chemical risk amplification factor, with a value of hazardous chemical level / 9; α, β, γ, and δ are weighting coefficients, satisfying α + β + γ + δ = 1.

8. The method according to claim 7, characterized in that, The weighting coefficients α, β, γ, and δ are dynamically switched according to the current warning level: During typhoon warning periods: α=0.5, β=0.2, γ=0.1, δ=0.2; During a rainstorm warning: α=0.1, β=0.6, γ=0.1, δ=0.2; During normal operation: α=0.2, β=0.2, γ=0.1, δ=0.

5.

9. The method according to claim 1, characterized in that, The preset safety threshold τ in step S6 is set according to the current terminal safety level: Level 1 safety level τ=0.3, Level 2 safety level τ=0.5, Level 3 safety level τ=0.7; and / or, the objective function of the optimization model is min C=ω_1×D(i)+ω_2×F(i), where D(i) is the horizontal transport distance of the candidate container from the gate or ship berth to the tank i, F(i) is the cost of reloading caused by being assigned to tank i, and ω_1 and ω_2 are weighting coefficients.

10. A dynamic safety location system for container terminal yards based on multimodal environmental perception and digital twin inference, characterized in that, include: The multimodal environment sensing layer is used to deploy a gridded multimodal environment sensor array in the yard to collect multidimensional environmental parameters at a preset period. The data fusion and preprocessing layer is used to receive the raw sensor data from the multimodal environment perception layer, perform timestamp synchronization, outlier filtering and Kalman filtering smoothing, and output a fused environment parameter matrix with a unified time resolution. The digital twin simulation engine is used to build and maintain a high-precision three-dimensional digital twin model of the storage yard, and to infer the spatial distribution of environmental risks within a future time window based on the fused environmental parameter matrix, generating a heat map of environmental risks in the storage yard. The container-environment coupling risk assessment module is used to extract attribute data for each candidate container to construct a coupling risk feature vector, and perform spatial matching and coupling calculation with the environmental risk heat map to obtain the container-environment coupling risk index CRI(i) for each candidate slot. The dynamic safety location selection decision module is used to construct a feasible set with CRI less than a preset safety threshold τ as a hard constraint, solve an optimization model within the feasible set with the goal of minimizing the operating cost, and output the optimal slot. An adaptive pre-evacuation rescheduling module is used to trigger an alarm and switch to the pre-evacuation process if the feasible set is empty.

11. The system according to claim 10, characterized in that, The adaptive pre-evacuation rescheduling module is also used to automatically trigger the pre-evacuation process when environmental parameters exceed the preset secondary warning threshold, identify high-risk container sets, and generate a redistribution scheme with the goal of minimizing the total evacuation cost.

12. The system according to claim 10, characterized in that, In the multimodal environmental perception layer, the sensors deployed in each grid node include: ultrasonic anemometer, tide gauge, rain gauge, visibility sensor and triaxial accelerometer seismic sensor; each sensor uploads data in real time through industrial Ethernet or LoRa wireless sensor network, with a data acquisition cycle of 5-10 minutes.

13. The system according to claim 10, characterized in that, The dynamic safety location selection decision module uses the ε-constraint method to construct an optimization model. The CRI being less than the preset safety threshold τ is an inviolable hard constraint. The optimization objective is to minimize the weighted sum of horizontal transport distance and container handling cost. The optimal slot is solved by an integer programming solver or a heuristic algorithm. The decision results are sent to the yard crane PLC and truck scheduling system for execution via MQTT or OPC UA protocol.

14. The system according to claim 11, characterized in that, In the adaptive pre-evacuation rescheduling module: the secondary early warning thresholds include wind speed threshold, rainfall threshold, tide height threshold, and earthquake acceleration threshold; the high-risk container set is the set of containers whose CRI exceeds the evacuation threshold τ_evac; the reassignment scheme takes minimizing the total evacuation cost as the objective function and constrains the newly assigned slots to be located within the safe area where CRI < τ; the system generates an evacuation instruction sequence according to the priority order of hazardous chemical containers first, loaded containers second, and empty containers last.

Images