A ship lock full-automatic lock passing decision method based on multi-factor analysis

Abstract

The present application relates to a kind of ship lock full automation lock decision-making methods based on multi-factor analysis, first, four big core influence factors such as tide, water conservancy intervention, rainfall, traffic demand are comprehensively covered, provide solid data support for lock decision-making;Then long short-term memory neural network is used to build Yangtze side and canal side water level prediction model, can effectively capture long-term dependence in time series data, finally based on reinforcement learning, an intelligent decision-making model for lock opening is constructed, realizing the full-process intelligence from feature recognition, data processing, water level prediction to opening and closing decision, which can dynamically adjust the decision according to real-time conditions, output optimal lock opening decision and time window. Realize the intelligence of lock decision-making, improve the efficiency and safety of ship lock.

Description

Technical Field

[0001] This invention relates to a fully automated lock opening decision-making method based on multi-factor analysis, belonging to the field of waterway engineering management, and is applicable to the intelligent management of locks affected by multiple factors such as natural conditions, water conservancy facilities, traffic demand, and operating hours. Background Technology

[0002] Lock-through mode is one of the core modes for efficient operation of ship locks. It refers to a mode where, under specific safety conditions, the upstream and downstream gates of the lock open simultaneously, allowing ships to pass through without needing to go through a water replenishment / discharge process within the lock chamber. Lock-through mode eliminates the need to open and close lock valves, significantly improving navigation efficiency and reducing energy consumption. This mode is currently mainly used in tidal river sections and river-sea confluence areas, optimizing lock throughput capacity while reducing energy consumption, making it an important operational solution for the efficient development of modern shipping.

[0003] Currently, the decision-making process for opening and closing ship locks in the Yangtze River basin relies primarily on manual experience. However, the changes in water level differences on both sides of the locks are influenced by a combination of multiple factors, exhibiting extreme dynamism and complexity. For example, the Yangtze River's tidal forces cause periodic fluctuations in water level, and the propagation of tidal levels to the lock area involves a phase delay effect. The operation of surrounding control gates and pumping stations directly alters the water level balance, with the impact being significantly stronger during the flood season than during the non-flood season. While rainfall conditions generally have a weaker impact, continuous torrential rains can trigger sudden changes in water level. Furthermore, fluctuations in traffic demand and restrictions on navigation hours further increase the difficulty of decision-making. Therefore, manual decision-making struggles to accurately grasp these uncertainties, fails to fully exploit the efficiency potential of lock opening modes, and may even pose safety hazards.

[0004] Existing lock decision-making systems are often designed for traditional lock-locking modes, and the factors considered are difficult to match the actual needs of lock operation, making them unsuitable for complex and ever-changing natural, engineering, and traffic environments. Furthermore, insufficient accuracy in predicting water levels and water level differences also affects the scientific validity of lock operation decisions. Therefore, to address the shortcomings of existing technologies, accurately predicting the water level difference changes on both sides of the lock based on the safety and efficiency requirements of lock operation modes, and achieving intelligent decision-making for lock opening and closing, has become a pressing technical problem in this field. Summary of the Invention

[0005] This invention provides a fully automated lock opening decision method based on multi-factor analysis, which realizes intelligent lock opening decision and improves lock throughput efficiency and passage safety.

[0006] The technical solution adopted by this invention to solve its technical problem is: A fully automated lock opening decision-making method based on multi-factor analysis includes the following steps: Step S1: Identify the characteristic factors affecting the change in water level difference on both sides of the lock, and determine the relevant influencing factors through the characteristic factors; The characteristic factors include tidal action, water conservancy facility intervention, rainfall conditions, and traffic demand. Tidal action is characterized by water level data at the Yangtze River estuary. Water conservancy facility intervention includes the drainage and diversion flow of the control gates around the lock, the pumping flow and operating power of the pumping station, rainfall conditions are reflected by rainfall data from nearby meteorological stations upstream and downstream, and traffic demand is measured by the total tonnage of ships waiting to pass through the lock at each decision point. Step S2: Collect data on influencing factors related to the feature factors in step S1, process the data on influencing factors, and use the random forest model to calculate and analyze the weight of each influencing factor. The process of processing the influencing factor data involves standardizing the selected raw data using Z-Score to transform it into effective features that conform to the logic of water level difference prediction. The random forest model quantifies the impact weight of each effective feature on the accuracy of water level difference prediction based on the variance reduction brought by the effective features at the split nodes. Step S3: Based on the influencing factors and their influence weights identified in step S2, construct a water level prediction model for the Yangtze River side of the lock and a water level prediction model for the canal side. Output water level prediction values ​​through the two models, and then obtain the predicted water level difference between the two sides of the lock. Step S4: Based on the predicted water level difference obtained in step S3, reinforcement learning is used to construct an intelligent gate opening and closing decision model for the gate operation mode. After comprehensively considering the characteristic factors, the gate opening decision and gate opening time window for the gate operation mode are given. Among them, reinforcement learning adopts the DQN algorithm, whose state space is the feature factors, action space is the gate opening and closing, and reward function comprehensively considers the lock passage efficiency and navigation safety, and performs autonomous learning and iterative calculation to finally obtain the optimal gate opening decision under different states. Furthermore, in step S1, the tidal effect causes a phase delay effect after the tide level propagates to the lock area. Therefore, it is necessary to obtain the correlation coefficient under different delays by using the water level data of the Yangtze River estuary. First, the Peer correlation coefficient between the shifted sequence and the original sequence is calculated. The calculation formula is as follows: (1); In formula (1), Lagging Correlation coefficient at time, x For tide level data, for Keep moving forward Hourly tide level y This refers to water level data on the Yangtze River side. y i For the first iThe water level value of the Yangtze River at that moment. and These are the average values ​​of the corresponding samples. n The number of samples; The optimal delay time is determined based on the correlation coefficients for different lag times, and the calculation formula is as follows: (2); In formula (2), Lagging Correlation coefficient at time, This is the maximum permissible delay time; The formula for calculating the correlation between water level difference and water conservancy facility intervention is as follows: (3); In formula (3), The correlation coefficient for water conservancy facility intervention is given, with a value range of [-1, 1]. For the sample size, Y is the manipulated variable, which includes pumping flow rate and net flow rate; and Y is the response variable, which includes water level difference. For the first Traffic value at each point in time, For the first Water level values ​​at each time point This represents the average flow rate. This represents the average water level. In the aforementioned rainfall conditions, water level change is used as a reflection of rainfall amount; therefore, the first... hourly water level change The calculation formula is: (4); In formula (4), for Change in water level over time for The actual measured absolute water level at any given time. This represents the measured absolute water level for the previous hour.

[0007] Furthermore, the specific steps in step S2 are as follows: Step S21: A manual feature selection strategy based on the physical mechanisms of hydraulic engineering is used to select 25-30 influencing factors as feature variables. Feature construction is achieved through time series shifting, transforming the variable values ​​predicted for future moments into features using historical values ​​and the current value as the label. The time series shifting method is as follows: (5); In formula (5), for tCharacteristic variables at time, The constructed feature vector includes past Observed values ​​of influencing factors at each time point; Step S22: Since the feature construction uses a time series shift method, the data header may be missing. Therefore, the feature cleaning method uses list deletion to remove missing data. The data list after feature cleaning is as follows: (6); In formula (6), This represents the source of the filtering, i.e., the initial dataset after formula (5) is completed. This represents the result set after cleaning. A subset of; Step S23: Perform feature transformation on the cleaned data, using Z-Score standardization. The processing formula is as follows: (7); In formula (7), These are the original feature values ​​after cleaning. This is the sample mean of this feature column. This is the sample standard deviation of this feature column. The transformed eigenvalues ​​follow a standard normal distribution with a mean of 0 and a variance of 1. Step S24, assume any intermediate node Includes Calculate the variance of water level difference data within a given historical observation sample. As an indicator for measuring the uncertainty of current water level difference prediction, its calculation formula is as follows: (8); In formula (8), For nodes The Middle The true water level difference of each sample For nodes The average water level difference across all samples; Step S25: Iterate through the feature values ​​transformed in step S23, select one feature value as the splitting feature, and select a specific water level value as the splitting threshold. Split into left child node and right child node The reduction in mean square error caused by the splitting feature at this node Calculate according to the following formula: (9); In formula (9), It is a characteristic of splitting. This represents the total fluctuation of water level difference data within a node before the introduction of splitting features. This represents the number of samples whose splitting feature is less than the splitting threshold. The number of samples whose splitting feature is greater than the splitting threshold. , These are the residual variances of the water level difference data within the two subsets after classification using the splitting feature; Step S26: Obtain the expected value of all contributions of the feature values ​​in the entire random forest model, i.e. (10); In formula (10), The total number of decision trees in the random forest. For the first The set of all nodes of the regression tree For indicator functions; Step S27, perform normalization operation, the formula is: (11); The formula (11) is used to calculate the result. The weight of the selected splitting feature on the water level difference change is the final weight of each influencing factor.

[0008] Furthermore, in step S3, the weights obtained in step S2 are used to select influencing factors within a set range according to the order, which are then used as input to the LSTM model. The last hidden state output is passed through a four-layer fully connected network to obtain the prediction result. Water level prediction models for the Yangtze River side and the canal side of the lock are constructed respectively. Water level prediction values ​​are output through the two models, thereby obtaining the predicted value of the water level difference on both sides of the lock. Among them, the fully connected network introduces the ReLU activation function to fit the complex water level change curve and introduces Dropout to prevent overfitting. Furthermore, the specific steps of step S4 are as follows: Step S41: Establish the state space, determine the model input, and the model output is the predicted water level difference obtained in step S3. Step S42, construct the action space at each moment. Select a discrete action : ; Step S43, design the reward function, including gate opening reward and gate closing reward; The gate opening reward It consists of two parts: hard constraints and soft benefits. (12) In (12), AThe preset value for hard constraint circuit breaker penalty is used to regulate illegal gate opening behavior and is set according to the lock safety level; H safe To preset a safe water level difference threshold, which characterizes the safe water level baseline for the operation of the ship lock; T start ,T end The preset navigation time intervals conform to the daily operation and scheduling rules of the lock; R safe To preset a safe rainfall threshold and avoid navigation risks caused by heavy rain; When all the above hard constraints are met, a soft-reward bonus is triggered. The calculation formula is as follows: (13); In formula (13), The penalty measures are designed to address the risk of future water level differences exceeding the standard, reflecting considerations for long-term navigation safety. To incentivize efficiency, decisions to open the gates when conditions for continuous and safe navigation are met are encouraged, thereby improving traffic efficiency; Tonnage-based incentives will be provided to ensure the passage of large-tonnage vessels and optimize the efficiency of shipping resource allocation. The future risk penalty in formula (13) is: (14) In formula (14), P risk_base Preset values ​​for risk-based penalties. h t+1 Forecast water level difference for the next hour. K risk This is a risk amplification factor, adjusted based on the hydrological sensitivity of the lock. d 1 represents the discount factor for the next hour, used to weigh near-term risk. The efficiency reward is: (15) In formula (15), P eff The preset value for the efficiency reward coefficient; d 2 represents the efficiency discount factor, adapting to the continuous navigation needs at different time scales; T safe For the number of consecutive safe flight hours in the future; T safe_min The minimum continuous safe navigation duration threshold is preset; P eff_penalty A preset value is used to penalize insufficient efficiency; The tonnage bonus is: (16) In formula (16), P ton Preset values ​​for navigation incentives for large-tonnage vessels; w t The total tonnage of ships waiting to enter the lock; W threshold A preset tonnage threshold is used to distinguish between large and small tonnage vessels; The gate closure reward is as follows: (17) In formula (17), The default value for the base penalty of missing a perfect air traffic window; To incur additional penalties for missing out on large-tonnage vessel passage, a tonnage threshold can be applied. W threshold Dynamic adjustment; P close_base The preset value for the basic reward for closing the gate; P avoid An additional reward is preset to encourage sluice gate closure as a safety precaution when water levels exceed the warning level. For A perfect navigation window requires five constraints simultaneously, namely (1) the current water level is safe, i.e., the absolute value of the current water level difference. H safe ； (2) Time period compliance, i.e., the current time period (3) Efficiency targets are met, i.e., the predicted continuous safe navigation time in the future is achieved. (4) The weather is suitable, that is, the current rainfall is suitable. r t ≤ R safe (5) Low future risk, i.e., the absolute value of the predicted water level difference in the next hour. H safe .

[0009] By employing the above technical solutions, the present invention has the following beneficial effects compared to the prior art: 1. The fully automated lock opening decision method based on multi-factor analysis provided by this invention comprehensively covers four core influencing factors: tides, water conservancy facility intervention, rainfall, and traffic demand. It realizes multi-factor systematic analysis of lock opening decisions and provides solid data support for lock opening decisions. 2. The fully automated lock opening decision method based on multi-factor analysis provided by this invention uses a long short-term memory neural network (LSTM) to construct a water level prediction model on the Yangtze River side and the canal side. It can effectively capture long-term dependencies in time series data, has high prediction accuracy, and provides a reliable reference for water level changes for lock opening decisions. 3. The fully automated lock opening decision method based on multi-factor analysis provided by this invention constructs an intelligent decision model for lock opening based on reinforcement learning, realizing full-process intelligence from feature identification, data processing, water level prediction to opening and closing decisions. It can dynamically adjust decisions according to real-time conditions, output the optimal lock opening decision and time window, improve the stability and intelligence level of lock operation, and reduce the operational risks caused by human decision-making bias. Attached Figure Description

[0010] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0011] Figure 1 This is the overall flowchart of the fully automated lock opening decision method based on multi-factor analysis provided by the present invention; Figure 2 This is a location map of Wusong Station used in the embodiments provided by the present invention; Figure 3 This is a diagram showing the relative positions of the control gates in the embodiments provided by the present invention; Figure 4 This is a location diagram of Dantu District and Danyang Station in the embodiments provided by the present invention; Figure 5 This is a structural diagram of the Yangtze River water level prediction LSTM model provided in the embodiments of the present invention; Figure 6 , Figure 7 This is a comparison of the actual value and the predicted value at different times in the embodiments provided by the present invention; Figure 8 This is a structural diagram of the LSTM model for canal water level prediction provided in the embodiments of the present invention; Figure 9 , Figure 10 , Figure 11 This is a diagram illustrating the predicted canal water level in an embodiment provided by the present invention. Figure 12 , Figure 13 , Figure 14 This is a diagram illustrating the water level difference prediction effect in an embodiment provided by the present invention; Figure 15 This is a framework diagram of the intelligent start-up and shutdown decision-making model provided in the embodiments of the present invention; Figure 16 , Figure 17 This is a training effect diagram of the intelligent start-stop decision model in the embodiment provided by the present invention. Detailed Implementation

[0012] The present invention will now be described in further detail with reference to the accompanying drawings. In the description of this application, it should be understood that the terms "left side," "right side," "upper part," "lower part," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. "First," "second," etc., do not indicate the importance of the components, and therefore should not be construed as a limitation of the present invention. The specific dimensions used in this embodiment are only for illustrating the technical solution and do not limit the scope of protection of the present invention.

[0013] As described in the background section, existing lock decision-making systems are mostly designed for traditional lock-locking modes, failing to fully adapt to the actual needs of lock operation. Furthermore, their accuracy in predicting water level and water level difference is insufficient, making it difficult to cope with complex and ever-changing natural, engineering, and traffic environments. Therefore, this application provides a fully automated lock operation decision-making method based on multi-factor analysis, balancing scientific rigor, safety, and efficiency.

[0014] like Figure 1 The diagram shown is a schematic of the overall process of the fully automated lock opening decision method based on multi-factor analysis provided in this application. It mainly includes four steps. The first step, S1, is to identify the characteristic factors affecting the change of water level difference on both sides of the lock and determine the relevant influencing factors through the characteristic factors. Among them, the characteristic factors include tidal action, water conservancy facility intervention, rainfall conditions, and traffic demand. Tidal action is represented by water level data of the Yangtze River estuary. Water conservancy facility intervention includes the drainage flow and water diversion flow of the control gates around the lock, the pumping flow and operating power of the pumping station, rainfall conditions are reflected by the rainfall data of nearby meteorological stations upstream and downstream, and traffic demand is measured by the total tonnage of ships waiting to pass through the lock at each decision point.

[0015] The tidal action causes a phase delay effect after the tide level propagates to the lock area. Therefore, it is necessary to characterize this effect using water level data from the Yangtze River estuary, and then analyze the phase delay effect of the tide level propagation to the lock area, obtaining the correlation coefficients under different delays. First, the Pierre correlation coefficient between the shifted sequence and the original sequence is calculated. For a specific delay... The calculation formula is: (1); In formula (1), Lagging Correlation coefficient at time, x For tide level data, for Keep moving forward Hourly tide level y This refers to water level data on the Yangtze River side.y i For the first i The water level value of the Yangtze River at that moment. and These are the average values ​​of the corresponding samples. n The number of samples; The optimal delay time is determined based on the correlation coefficients for different lag times, and the calculation formula is as follows: (2); In formula (2), Lagging Correlation coefficient at time, This is the maximum permissible delay time; The formula for calculating the correlation between water level difference and water conservancy facility intervention is as follows: (3); In formula (3), The correlation coefficient for water conservancy facility intervention is given, with a value range of [-1, 1]. For the sample size, Y is the manipulated variable, which includes pumping flow rate and net flow rate; and Y is the response variable, which includes water level difference. For the first Traffic value at each point in time, For the first Water level values ​​at each time point This represents the average flow rate. This represents the average water level. In the aforementioned rainfall conditions, water level change is used as a reflection of rainfall amount; therefore, the first... hourly water level change The calculation formula is: (4); In formula (4), for Change in water level over time for The actual measured absolute water level at any given time. This represents the measured absolute water level for the previous hour.

[0016] Continuing with step S2, this step is mainly to calculate the weights of influencing factors. However, in order to improve the accuracy of the data, the data is processed first. Through the process of "feature selection → feature construction → feature cleaning → feature transformation", the raw data is transformed into effective features that conform to the water level difference prediction logic.

[0017] In the feature selection stage, an artificial feature selection strategy based on the physical mechanisms of hydraulic engineering was adopted. Ultimately, a certain number of feature variables with clear physical significance were selected. Based on practical experience, for example, in terms of physical inertia, the water level sequence of the Yangtze River and the Grand Canal with a lag of 1 to 12 hours was selected according to the continuous physical characteristics of water level changes; in terms of physical delay, combined with the tidal transmission time pattern revealed by the previous lag analysis, only the tidal level data from 5 to 8 hours ago was retained; in terms of instantaneous impact, the immediate effect of human regulation was considered, and the real-time flow data of the control gate and pumping station at the current moment was included.

[0018] The specific operation process described above is as follows: Step S21, using an artificial feature selection strategy based on the physical mechanism of hydraulic engineering to select 25-30 influencing factors as feature variables. Feature construction is achieved through time series shifting, transforming the variable values ​​predicted for future moments into variable values ​​at historical moments as features, and the current moment value as the label. The time series shifting method is as follows: (5); In formula (5), for t Characteristic variables at time, The constructed feature vector includes past Observed values ​​of influencing factors at each time point; Step S22: Since the feature construction uses a time series shift method, the data header may be missing. Therefore, the feature cleaning method uses list deletion to remove missing data. The data list after feature cleaning is as follows: (6); In formula (6), This represents the source of the filtering, i.e., the initial dataset after formula (5) is completed. This represents the result set after cleaning. A subset of; Step S23: Perform feature transformation on the cleaned data, using Z-Score standardization. The processing formula is as follows: (7); In formula (7), These are the original feature values ​​after cleaning. This is the sample mean of this feature column. This is the sample standard deviation of this feature column. The transformed eigenvalues ​​follow a standard normal distribution with a mean of 0 and a variance of 1. To determine the model's dependence on a particular feature during prediction, and considering the interactions between features, their practical value in the model, and potential nonlinear relationships, a random forest regression model consisting of several decision trees is constructed. The importance of a feature is quantified by the variance reduction it causes at split nodes. Within the random forest regression model, the system uses a perturbation test method to quantify the weight of each input feature's influence on the prediction accuracy of the target variable (water level difference).

[0019] That is, in step S24, during the construction of the regression tree, it is assumed that any intermediate node... Includes Calculate the variance of water level difference data within a given historical observation sample. As an indicator for measuring the uncertainty of current water level difference prediction, its calculation formula is as follows: (8); In formula (8), For nodes The Middle The true water level difference of each sample For nodes The average water level difference across all samples; Step S25: Iterate through the feature values ​​transformed in step S23, select one feature value as the splitting feature, and select a specific water level value as the splitting threshold. Split into left child node and right child node The reduction in mean square error caused by the splitting feature at this node Calculate according to the following formula: (9); In formula (9), It is a characteristic of splitting. This represents the total fluctuation of water level difference data within a node before the introduction of splitting features. This represents the number of samples whose splitting feature is less than the splitting threshold. The number of samples whose splitting feature is greater than the splitting threshold. , These are the residual variances of the water level difference data within the two subsets after classification using the splitting feature; Step S26: Obtain the expected value of all contributions of the feature values ​​in the entire random forest model, i.e. (10); In formula (10), The total number of decision trees in the random forest. For the first The set of all nodes of the regression tree For indicator functions; Step S27, perform normalization operation, the formula is: (11); The formula (11) is used to calculate the result. The weight of the selected splitting feature on the change in water level difference is the final weight of each influencing factor (specifically, the weight of the Yangtze River's water level 1 hour ago on the change in water level difference).

[0020] Step S3: Based on the influencing factors and their weights identified in Step S2, construct a water level prediction model for the Yangtze River side and a water level prediction model for the canal side of the lock. The two models output predicted water levels, thus obtaining the predicted water level difference between the two sides of the lock. Here, the model is based on an LSTM model. Influencing factors within a set range are selected according to their order and used as input to the LSTM model. The last hidden state output is processed by a four-layer fully connected network to obtain the prediction result. The two models are then used to construct the water level prediction model for the Yangtze River side and the water level prediction model for the canal side of the lock. The fully connected network uses the ReLU activation function to fit complex water level change curves and introduces Dropout to prevent overfitting.

[0021] Taking the aforementioned preferred influencing factors as an example, the inputs of the Yangtze River side water level prediction model for the lock include the tide level data of the Yangtze River estuary 4, 5, 6, and 7 hours ago, the historical water level values ​​of the past 12 hours, and the flow data of the pumping station and the control gate. The prediction output is the water level value of the Yangtze River side of the lock for the next three hours. The inputs of the canal side water level prediction model include the historical water levels of the Yangtze River and the canal for the past 12 hours, the real-time flow of the current control gate and pumping station, and the Yangtze River tide levels 5, 6, 7, and 8 hours ago. The prediction output is the water level value of the canal side of the lock for the next three hours, and the water level difference data between the upstream and downstream of the lock for the next three hours is calculated based on the above prediction results.

[0022] The specific input-output method is as follows: LSTM is the core of the model, used to solve the gradient vanishing problem in long sequence training. For each time step t in the input sequence, the LSTM unit performs the following calculations: (1) Forget Gate - Historical Water Information Screening Mechanism: ; in, The input feature vector for the current water level prediction (30-dimensional for the canal side, 18-dimensional for the Yangtze River side). The hidden state of the previous moment: refers to the short-term trend of water level changes at time t-1; Forgetting coefficient: refers to the correlation weight of historical water conditions; the model uses this to determine the water level trend at the previous moment. Whether the long-term accumulated water volume is still valid in the present situation.

[0023] (2) Input Gate - New Water Information Integration Mechanism: ; ; in, Input gating: refers to the degree to which current hydrological information is accepted. It determines the extent to which a surge in tidal level or the activation of water pumps will alter the system's internal memory. Candidate memory state: refers to a new water level feature derived from the current observation; , The weight matrix is ​​responsible for mapping the input low-dimensional hydrological physical quantities to a high-dimensional feature space to fit the complex nonlinear functional relationships in the process of water level change.

[0024] (3) Cell State Update - Core Water Level Memory Transmission: ; in, The cell state at the previous moment: refers to the long-term hydrological memory of the system, which is not just the water level of the previous hour, but includes the background water level or long-term tidal effects accumulated over the past 12 hours. The current cell state refers to the updated long-term hydrological memory characteristics of the canal system, which consists of two parts: one part is the retained baseline water level from the previous period (…). The other part is the increase in water level caused by the current tides and regulation. ).

[0025] (4) Output Gate - Predictive Feature Generation: ; ; in, Output gating: refers to a feature filter that, based on the current hydrological input, determines which information to extract as explicit features from the complex internal water state. The hidden state at the current time: refers to the high-dimensional water level feature vector processed at time t; It is not a direct prediction of the water level, but an abstract vector containing rich temporal information. It is passed into the subsequent fully connected layer and finally mapped to the specific water level values ​​of the canal in the 1st, 2nd and 3rd hours in the future.

[0026] Taking canal side water level prediction as an example, its input feature vector is composed as follows: ; in, This is the water level on the Yangtze River side over the past 12 hours. This represents the water level on the canal side over the past 12 hours. Delayed tide characteristics, i.e., tide levels at t-5, t-6, t-7, and t-8 hours. This represents the current flow rate of the control gate and the flow rate of the pumping station.

[0027] Taking canal-side water level prediction as an example again, the last hidden state output by the LSTM is passed through a 4-layer fully connected network to obtain the final prediction result. .

[0028] ; ; ; ; Among them, the introduction To enable the model to fit complex water level change curves, Dropout is introduced to randomly set the output of some neurons to zero during training, thus preventing overfitting. The output dimension is 3, corresponding to the predicted water level for the next 1 hour, 2 hours, and 3 hours respectively.

[0029] Finally, in step S4, based on the predicted water level difference obtained in step S3, an intelligent gate opening and closing decision model for the gate operation mode is constructed using reinforcement learning. After comprehensively considering characteristic factors, the gate opening decision and gate opening time window for the gate operation mode are given.

[0030] Specifically, step S41 is to establish the state space, determine the model input, and the model output is the predicted water level difference obtained in step S3. Step S42, construct the action space at each moment. Select a discrete action : ; Step S43, design the reward function, including gate opening reward and gate closing reward; The gate opening reward It consists of two parts: hard constraints and soft benefits. (12) In (12), A The preset value for hard constraint circuit breaker penalty is used to regulate illegal gate opening behavior and is set according to the lock safety level;H safe To preset a safe water level difference threshold, which characterizes the safe water level baseline for the operation of the ship lock; T start ,T end The preset navigation time intervals conform to the daily operation and scheduling rules of the lock; R safe To preset a safe rainfall threshold and avoid navigation risks caused by heavy rain; When all the above hard constraints are met, a soft-reward bonus is triggered. The calculation formula is as follows: (13) In formula (13), The penalty measures are designed to address the risk of future water level differences exceeding the standard, reflecting considerations for long-term navigation safety. To incentivize efficiency, decisions to open the gates when conditions for continuous and safe navigation are met are encouraged, thereby improving traffic efficiency; Tonnage-based incentives will be provided to ensure the passage of large-tonnage vessels and optimize the efficiency of shipping resource allocation. The future risk penalty in formula (13) is: (14) In formula (14), P risk_base Preset values ​​for risk-based penalties. h t+1 Forecast water level difference for the next hour. K risk This is a risk amplification factor, adjusted based on the hydrological sensitivity of the lock. d 1 represents the discount factor for the next hour, used to weigh near-term risk. The efficiency reward is: (15) In formula (15), P eff The preset value for the efficiency reward coefficient; d 2 represents the efficiency discount factor, adapting to the continuous navigation needs at different time scales; T safe For the number of consecutive safe flight hours in the future; T safe_min The minimum continuous safe navigation duration threshold is preset; P eff_penalty A preset value is used to penalize insufficient efficiency; The tonnage bonus is: (16) In formula (16),P ton Preset values ​​for navigation incentives for large-tonnage vessels; w t The total tonnage of ships waiting to enter the lock; W threshold A preset tonnage threshold is used to distinguish between large and small tonnage vessels; The gate closure reward is as follows: (17) In formula (17), The default value for the base penalty of missing a perfect air traffic window; To incur additional penalties for missing out on large-tonnage vessel passage, a tonnage threshold can be applied. W threshold Dynamic adjustment; P close_base The preset value for the basic reward for closing the gate; P avoid An additional reward is preset to encourage sluice gate closure as a safety precaution when water levels exceed the warning level. For A perfect navigation window requires five constraints simultaneously, namely (1) the current water level is safe, i.e., the absolute value of the current water level difference. H safe ； (2) Time period compliance, i.e., the current time period (3) Efficiency targets are met, i.e., the predicted continuous safe navigation time in the future is achieved. (4) The weather is suitable, that is, the current rainfall is suitable. r t ≤ R safe (5) Low future risk, i.e., the absolute value of the predicted water level difference in the next hour. H safe .

[0031] In summary, this application 1. Constructs a multi-source feature identification and data processing system: First, it identifies four core feature factors affecting the water level difference on both sides of the lock, including the Yangtze River estuary water level representing tidal effects, the flow and power of surrounding control gates and pumping stations representing water conservancy facility intervention, the upstream and downstream rainfall representing meteorological conditions, and the total tonnage of waiting vessels representing traffic demand; the above data are collected and Z-Score standardized, and the variance reduction of each factor at the split node is calculated using a random forest model, quantifying and screening out key effective features that have a significant impact on water level changes.

[0032] 2. Constructing a two-sided independent water level prediction model based on Long Short-Term Memory (LSTM) network: Based on the selected key features, separate water level prediction models are established for the Yangtze River side and the canal side. The model input includes the historical 12-hour water level sequence and the delayed tide features from 5-8 hours ago. The unique gating mechanism of the LSTM network is used to capture the long-sequence time dependence and tide lag effect, and the independent water level prediction values ​​for future times are output respectively. Then, a high-precision prediction sequence of water level difference on both sides of the lock is calculated, which serves as the physical environment input of the decision system.

[0033] 3. Construct an intelligent gate opening and closing decision model based on deep reinforcement learning (DQN): Using feature factors and predicted water level difference as the state space and gate opening and closing as the action space, design a two-layer reward function that integrates "hard constraint circuit breaking" (water level / time / rainfall red line) and "multi-dimensional benefit evaluation" (large ship reward / continuous navigation incentive); through the interaction and iteration between the DQN agent and the environment, the prediction error is minimized by using Smooth L1 Loss, and the globally optimal strategy under different hydrological conditions is learned. Finally, the optimal gate opening decision and gate opening time window that meet the dual constraints of safety and efficiency are output.

[0034] To verify the feasibility of the above method, this application provides a specific embodiment.

[0035] The first step involved selecting the Jianbi Ship Lock. Through on-site investigation and data analysis, the characteristic factors influencing the water level difference on both sides of the lock were identified, including tidal forces, intervention by hydraulic facilities (operation of the control gate and pumping station), rainfall conditions, season, and weather. Among these, tidal forces primarily affect the periodic changes in the Yangtze River water level; the control gate, used for flood control, irrigation, and navigation water level regulation, alters the water level during operation; pumping stations play a role in flood drainage and drought relief, similarly affecting the water level; rainfall may also have some impact on the water level. Analysis of the correlation between these influencing factors led to the following conclusions: Tidal effect: Considering the phase delay phenomenon in the propagation of tide level to the lock area, this study found that the optimal phase delay of tidal effect is 6 hours by quantifying the correlation under different lag times.

[0036] Intervention in water conservancy facilities: During the flood season from April to September each year, the control gates and pumping stations have a greater impact on water levels, with an impact intensity approximately 1.5 times that of the non-flood season.

[0037] Rainfall: On both the Yangtze River and the Grand Canal sides, the correlation coefficient between cumulative rainfall and water level changes in the first four hours was relatively high, but still below the significance level. Therefore, except in cases of severe and continuous torrential rain, the impact of rainfall on water level differences is relatively small.

[0038] The second step is to collect relevant data from 2022 to 2024, including: Upstream and downstream water levels and gate data: Complete records of water level changes and information at different times over three years.

[0039] Tide level information: Water level data from the Wusong station at the Yangtze River estuary, used to analyze the impact of tides on the Yangtze River water level, such as... Figure 2 The image shows the location of Wusong Station.

[0040] Operating data of control gates and pumping stations: including drainage flow and water intake flow of control gates, pumping flow and power of pumping stations, etc. Figure 3 The diagram shows the location of the control gate.

[0041] Rainfall data: Rainfall data from the meteorological station in Dantu District on the upper reaches of the Yangtze River and the Danyang station on the upper reaches of the Grand Canal at Jianbi Ship Lock. The locations of Dantu District and Danyang station are shown below. Figure 4 As shown.

[0042] Tonnage data: The tonnage data of vessels waiting to pass through the Jianbi Lock upstream and downstream can be used to characterize traffic demand and calculate traffic efficiency.

[0043] The collected data was processed first, including data cleaning to remove outliers and missing values. Then, feature engineering was performed to construct features related to water level difference prediction, such as water level and tide characteristics at different time lags. A random forest model was used to calculate the importance of each feature. After calculation and analysis, the results showed that the top fifteen most important influencing factors and their importance parameters were: ① Yangtze River water level 1 hour ago, ② Yangtze River water level 12 hours ago, ③ Canal water level 1 hour ago, ④ Wusongkou tide level 6 hours ago, ⑤ Canal water level 2 hours ago, ⑥ Canal water level 11 hours ago, ⑦ Wusongkou tide level 5 hours ago, ⑧ Canal water level 12 hours ago, ⑨ Yangtze River water level 11 hours ago, ⑩ Current sluice gate flow rate, ⑪ Yangtze River water level 10 hours ago, ⑫ Canal water level 3 hours ago, ⑬ Yangtze River water level 2 hours ago, ⑭ Wusongkou tide level 8 hours ago, and ⑮ Current pumping station flow rate.

[0044] The third step is the prediction model for the Yangtze River side water level of the ship lock: an LSTM model is used, the structure of which is as follows. Figure 5 As shown. Inputs include tide level data from Wusong Station 4, 5, 6, and 7 hours ago, Yangtze River water level data for the past 12 hours, and flow data from control gates and pumping stations. After model training, it predicts the Yangtze River water level for the next 1-3 hours, as shown. Figure 6 (Including 6a after 1 hour, 6b after 2 hours, and 6c after 3 hours) Figure 7The comparison between the actual and predicted values ​​is shown in Figures 7a (representing values ​​after 1 hour), 7b (representing values ​​after 2 hours), and 7c (representing values ​​after 3 hours). Hourly evaluation results of the water level prediction model show that the model performs best when the prediction duration is 1 hour, with a mean squared error (MSE) of 0.013019, a mean absolute error (MAE) of 0.082550, and a coefficient of determination (R²) of 0.993851. As the prediction duration increases to 2 hours, the MSE is 0.018339, the MAE is 0.095430, and the R² is 0.991339, maintaining high explanatory power. For predictions up to 3 hours, the MSE is 0.026664, the MAE is 0.114891, and the R² is 0.987410. This demonstrates that the Yangtze River water level prediction model for the lock possesses excellent time-series prediction capabilities, providing stable and reliable data support for water level reference needs at different time scales in the decision-making process for lock opening.

[0045] Canal side water level prediction model: Also using the LSTM model, the framework is as follows Figure 8 As shown. Inputs include the historical water levels of the Yangtze River and the Grand Canal over the past 12 hours, the real-time flow rates of the current control gates and pumping stations, and the tide levels from 5, 6, 7, and 8 hours ago. The model predicts the Grand Canal water level for the next 1-3 hours, and the comparison between the actual and predicted values ​​is shown below. Figure 9 (Including 9a after 1 hour, 9b after 2 hours, and 9c after 3 hours) Figure 10 (Including MSE comparison at time step 10a, MAE comparison at time step 10b, R² comparison at each time step 10c, and the training process at time step 10d) Figure 11 (Including 11a after 1 hour, 11b after 2 hours, and 11c after 3 hours) as shown. The hourly evaluation results of the water level prediction model show that the model performs best when the prediction duration is 1 hour, with an MSE of 0.0195, a MAE of 0.0797, and an R² of 0.8899. As the prediction duration increases to 2 hours, the MSE is 0.0249, the MAE is 0.0944, and the R² is 0.8543, maintaining high explanatory power. For 3-hour predictions: the MSE is 0.0313, the MAE is 0.1104, and the R² is 0.8210. It is evident that the model possesses stable and reliable time-series prediction performance. Considering the indirect influence of canal water levels on the control gates, pumping station regulation, and tide levels, the model can still accurately capture its changing patterns, providing strong data support for water level prediction at different time scales on the canal side in the decision-making process for opening gates.

[0046] Based on the predicted water levels on the Yangtze River side and the canal side, the predicted water level difference was further calculated as follows: Figure 12 , Figure 13 , Figure 14 As shown, the water level difference prediction result is used as the decision reference data for the gate opening and closing decision model in the fourth step.

[0047] The fourth step involves constructing an intelligent gate opening and closing decision-making model based on reinforcement learning. The principle and model framework are as follows: Figure 15 As shown. Based on actual scheduling requirements, the model's state space includes the current water level difference, the tonnage of vessels waiting to pass through the lock, the navigation period, and the predicted future water level difference; the action space is opening or closing the lock; the reward function comprehensively considers both traffic efficiency and safe passage objectives, i.e., positive rewards are given for improving traffic efficiency, and negative rewards are given for the occurrence of safety hazards.

[0048] In the Jianbi Lock's lock opening decision-making system, to achieve a dynamic balance between safety and navigation efficiency in lock scheduling, the mechanism is based on strict "red-line circuit breakers" and meticulous "benefit assessments": For lock opening decisions, the system sets three safety red lines. If the absolute value of the current water level difference exceeds 0.3 meters, it falls outside the non-navigable period of 7:00-19:00, or rainfall exceeds 20mm, a severe circuit breaker penalty of -50 points is imposed. Under compliance conditions, the model will accumulate scores based on four dimensions: the tonnage of vessels waiting to pass through the lock (a high reward of +25 points is given for vessels exceeding 20,000 tons), continuous efficiency (a reward of +12.9 points is given if the future continuous safe time is ≥2 hours). (Points awarded for failure to meet the standard, and a penalty of -5 points for failure to meet the standard) and future predicted risks (if the predicted water level exceeds the standard in the next hour, a dynamic heavy penalty will be imposed based on the base penalty of -5 points and combined with the amount of exceedance and coefficient); for the decision to close the gate, a base reward of +2 points is given by default. If the current water level is already in a dangerous state (>0.3m), an additional +5 points will be awarded (total +7 points) to encourage risk avoidance. However, if the gate is closed incorrectly under the "perfect navigation window" (i.e., the water level is safe, the time is appropriate, the efficiency meets the standard, there is no rainstorm, and the water level difference in the next hour is ≤0.3m), a penalty of -30 points for "missing an opportunity" will be imposed. If the tonnage of the waiting vessel that is refused entry at this time is >20,000 tons, the penalty will be further increased to -45 points.

[0049] Meanwhile, the Jianbi Lock water level intelligent decision-making system achieves dynamic optimization through a multi-timescale coupled prediction framework: integrating 1-hour, 2-hour, and 3-hour water level prediction models, and constructing a confidence-driven weighted fusion mechanism based on their actual prediction accuracy; under this framework, the system analyzes the evolution trend of water level difference in real time and generates time-varying decision strategies. This architecture significantly improves the robustness and timeliness of lock opening decisions through a closed-loop logic of multi-scale prediction → accuracy weighting → trend diagnosis → risk assessment.

[0050] Based on the analysis of influencing factors in the first and second steps and the calculation of their weights, a hierarchical risk assessment framework for sluice gate opening decisions is constructed. This framework sets the current water level difference, rainfall, and navigation period as the "core safety red line" (determining the feasibility of the operation), and sets the water level trend for the next hour and the duration of the continuous safety window as "auxiliary assessment indicators" (determining the stability and efficiency of the operation).

[0051] At the same time, the risk level is quantitatively defined: Risk-free (optimal decision): Both the core red line and auxiliary indicators are met. That is, the current environment is safe, the future water situation is stable, and there is a long-term navigation window.

[0052] Low risk (second-best efficiency): The core red lines are met, but the continuous safety window is insufficient. Although navigation is currently possible, the time window is short, and there is a risk of low operational efficiency.

[0053] Medium risk (trend warning): The core red line is met, but the water level is expected to exceed the standard in the next hour. This indicates that although the current situation is compliant, there is a potential risk of sharp fluctuations in the water level in the short term.

[0054] High Risk (Circuit Breaker Prohibited): Any core safety red line is not met. At this point, the environment is in an absolutely dangerous or non-compliant state, and the circuit breaker mechanism must be implemented immediately.

[0055] The DQN algorithm in reinforcement learning is used to solve the intelligent gate opening and closing decision model in the gate opening mode. The specific parameters are set as follows: 100 training rounds, batch size of 128 to improve GPU utilization, learning rate of 0.001, discount factor of 0.95, exploration rate linearly decaying with the total number of episodes (from 1.0 to 0.01), and experience replay buffer capacity of 100,000 records.

[0056] Assessment data shows that Figure 16 As shown, the vast majority of gate opening decisions were risk-free or low-risk, accounting for 93.1%, while medium-risk decisions accounted for only 6.9%, indicating that the system has significant reliability in ensuring gate opening security and operational compliance. The model's conditions for each condition being met during gate opening are as follows: Figure 17 As shown in the figure. According to statistics, 57.1% of the vessels in the lock-open mode are large vessels with a tonnage of ≥20,000 tons, which has a high passage efficiency.

[0057] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for fully automated lock opening decision based on multi-factor analysis, characterized in that: Includes the following steps: Step S1: Identify the characteristic factors affecting the change in water level difference on both sides of the lock, and determine the relevant influencing factors through the characteristic factors; The characteristic factors include tidal action, water conservancy facility intervention, rainfall conditions, and traffic demand. Tidal action is characterized by water level data at the Yangtze River estuary. Water conservancy facility intervention includes the drainage and diversion flow of the control gates around the lock, the pumping flow and operating power of the pumping station, rainfall conditions are reflected by rainfall data from nearby meteorological stations upstream and downstream, and traffic demand is measured by the total tonnage of ships waiting to pass through the lock at each decision point. Step S2: Collect data on influencing factors related to the feature factors in step S1, process the data on influencing factors, and use the random forest model to calculate and analyze the weight of each influencing factor. The process of processing the influencing factor data involves standardizing the selected raw data using Z-Score to transform it into effective features that conform to the logic of water level difference prediction. The random forest model quantifies the impact weight of each effective feature on the accuracy of water level difference prediction based on the variance reduction brought by the effective features at the split nodes. Step S3: Based on the influencing factors and their influence weights identified in step S2, construct a water level prediction model for the Yangtze River side of the lock and a water level prediction model for the canal side. Output water level prediction values ​​through the two models, and then obtain the predicted water level difference between the two sides of the lock. Step S4: Based on the predicted water level difference obtained in step S3, reinforcement learning is used to construct an intelligent gate opening and closing decision model for the gate operation mode. After comprehensively considering the characteristic factors, the gate opening decision and gate opening time window for the gate operation mode are given. The reinforcement learning uses the DQN algorithm, whose state space consists of feature factors and whose action space consists of opening and closing the gate. The reward function comprehensively considers the gate's passage efficiency and navigation safety, and performs autonomous learning and iterative calculation to finally obtain the optimal gate opening decision under different states.

2. The fully automated lock opening decision method based on multi-factor analysis according to claim 1, characterized in that: In step S1, the tidal effect causes a phase delay effect after the tide level propagates to the lock area. Therefore, it is necessary to obtain the correlation coefficient under different delays by using water level data from the Yangtze River estuary. First, the Pierre correlation coefficient between the shifted sequence and the original sequence is calculated. The calculation formula is as follows: （1）; In formula (1), Lagging Correlation coefficient at time, x For tide level data, for Keep moving forward Hourly tide level y This refers to water level data on the Yangtze River side. y i For the first i The water level value of the Yangtze River at that moment. and These are the average values ​​of the corresponding samples. n The number of samples; The optimal delay time is determined based on the correlation coefficients for different lag times, and the calculation formula is as follows: （2）; In formula (2), Lagging Correlation coefficient at time, This is the maximum permissible delay time; The formula for calculating the correlation between water level difference and water conservancy facility intervention is as follows: （3）; In formula (3), The correlation coefficient for water conservancy facility intervention is given, with a value range of [-1, 1]. For the sample size, Y is the manipulated variable, which includes pumping flow rate and net flow rate; and Y is the response variable, which includes water level difference. For the first Traffic value at each point in time, For the first Water level values ​​at each time point This represents the average flow rate. This represents the average water level. In the aforementioned rainfall conditions, water level change is used as a reflection of rainfall amount; therefore, the first... hourly water level change The calculation formula is: （4）; In formula (4), for Change in water level over time for The actual measured absolute water level at any given time. This represents the measured absolute water level for the previous hour.

3. The fully automated lock opening decision method based on multi-factor analysis according to claim 1, characterized in that: The specific steps in step S2 are as follows: Step S21: A manual feature selection strategy based on the physical mechanisms of hydraulic engineering is used to select 25-30 influencing factors as feature variables. Feature construction is achieved through time series shifting, transforming the variable values ​​predicted for future moments into features using historical values ​​and the current value as the label. The time series shifting method is as follows: （5）; In formula (5), for t Characteristic variables at time, The constructed feature vector includes past Observed values ​​of influencing factors at each time point; Step S22: Since the feature construction uses a time series shift method, the data header may be missing. Therefore, the feature cleaning method uses list deletion to remove missing data. The data list after feature cleaning is as follows: （6）; In formula (6), This represents the source of the filtering, i.e., the initial dataset after formula (5) is completed. This represents the result set after cleaning. A subset of; Step S23: Perform feature transformation on the cleaned data, using Z-Score standardization. The processing formula is as follows: （7）; In formula (7), These are the original feature values ​​after cleaning. This is the sample mean of this feature column. This is the sample standard deviation of this feature column. The transformed eigenvalues ​​follow a standard normal distribution with a mean of 0 and a variance of 1. Step S24, assume any intermediate node Includes Calculate the variance of water level difference data within a given historical observation sample. As an indicator for measuring the uncertainty of current water level difference prediction, its calculation formula is as follows: （8）; In formula (8), For nodes The Middle The true water level difference of each sample For nodes The average water level difference across all samples; Step S25: Iterate through the feature values ​​transformed in step S23, select one feature value as the splitting feature, and select a specific water level value as the splitting threshold. Split into left child node and right child node The reduction in mean square error caused by the splitting feature at this node Calculate according to the following formula: （9）; In formula (9), It is a characteristic of splitting. This represents the total fluctuation of water level difference data within a node before the introduction of splitting features. This represents the number of samples whose splitting feature is less than the splitting threshold. The number of samples whose splitting feature is greater than the splitting threshold. , These are the residual variances of the water level difference data within the two subsets after classification using the splitting feature; Step S26: Obtain the expected value of all contributions of the feature values ​​in the entire random forest model, i.e. （10）; In formula (10), The total number of decision trees in the random forest. For the first The set of all nodes of the regression tree For indicator functions; Step S27, perform normalization operation, the formula is: （11）; The formula (11) is used to calculate the result. The weight of the selected splitting feature on the water level difference change is the final weight of each influencing factor.

4. The fully automated lock opening decision method based on multi-factor analysis according to claim 1, characterized in that: In step S3, the weights obtained in step S2 are used to select influencing factors within a set range according to the order, which are then used as input to the LSTM model. The last hidden state of the output is passed through a four-layer fully connected network to obtain the prediction result. Water level prediction models for the Yangtze River side and the canal side of the lock are constructed respectively. Water level prediction values ​​are output through the two models, and then the water level difference prediction value on both sides of the lock is obtained. Among them, the fully connected network introduces the ReLU activation function to fit the complex water level change curve and introduces Dropout to prevent overfitting.

5. The fully automated lock opening decision method based on multi-factor analysis according to claim 1, characterized in that: The specific steps of step S4 are as follows: Step S41: Establish the state space, determine the model input, and the model output is the predicted water level difference obtained in step S3. Step S42, construct the action space at each moment. Select a discrete action : ; Step S43, design the reward function, including gate opening reward and gate closing reward; The gate opening reward It consists of two parts: hard constraints and soft benefits. （12） In (12), A The preset value for hard constraint circuit breaker penalty is used to regulate illegal gate opening behavior and is set according to the lock safety level; H safe To preset a safe water level difference threshold, which characterizes the safe water level baseline for the operation of the ship lock; T start ,T end The preset navigation time intervals conform to the daily operation and scheduling rules of the lock; R safe To preset a safe rainfall threshold and avoid navigation risks caused by heavy rain; When all the above hard constraints are met, a soft-reward bonus is triggered. The calculation formula is as follows: （13）; In formula (13), The penalty measures are designed to address the risk of future water level differences exceeding the standard, reflecting considerations for long-term navigation safety. To incentivize efficiency, decisions to open the gates when conditions for continuous and safe navigation are met are encouraged, thereby improving traffic efficiency; Tonnage-based incentives will be provided to ensure the passage of large-tonnage vessels and optimize the efficiency of shipping resource allocation. The future risk penalty in formula (13) is: （14） In formula (14), P risk_base Preset values ​​for risk-based penalties. h t+1 Forecast water level difference for the next hour. K risk This is a risk amplification factor, adjusted based on the hydrological sensitivity of the lock. d 1 represents the discount factor for the next hour, used to weigh near-term risk. The efficiency reward is: （15） In formula (15), P eff The preset value for the efficiency reward coefficient; d 2 represents the efficiency discount factor, adapting to the continuous navigation needs at different time scales; T safe For the number of consecutive safe flight hours in the future; T safe_min The minimum continuous safe navigation duration threshold is preset; P eff_penalty A preset value is used to penalize insufficient efficiency; The tonnage bonus is: （16） In formula (16), P ton Preset values ​​for navigation incentives for large-tonnage vessels; w t The total tonnage of ships waiting to enter the lock; W threshold A preset tonnage threshold is used to distinguish between large and small tonnage vessels; The gate closure reward is as follows: （17） In formula (17), The default value for the base penalty of missing a perfect air traffic window; To incur additional penalties for missing out on large-tonnage vessel passage, a tonnage threshold can be applied. W threshold Dynamic adjustment; P close_base The preset value for the basic reward for closing the gate; P avoid An additional reward is preset to encourage sluice gate closure as a safety precaution when water levels exceed the warning level. for A perfect navigation window requires five constraints simultaneously, namely (1) the current water level is safe, i.e., the absolute value of the current water level difference. H safe ； (2) Time period compliance, i.e., the current time period (3) Efficiency targets are met, i.e., the predicted continuous safe navigation time in the future is achieved. (4) The weather is suitable, that is, the current rainfall is suitable. r t ≤ R safe (5) Low future risk, i.e., the absolute value of the predicted water level difference in the next hour. H safe .

Images