Multimodal transport intelligent scheduling optimization method, apparatus and device, and storage medium

Abstract

The invention relates to the technical field of intelligent scheduling, and discloses a multimodal transport intelligent scheduling optimization method, device and equipment and a storage medium, and the method comprises the steps: carrying out the modeling of a plurality of transport nodes and transport paths, and obtaining a network topology structure; performing time period division on the scheduling period based on the network topology structure to obtain a time window scheduling framework; inputting the historical freight data into a multi-layer feedforward neural network for training to obtain a demand predictor; dynamically adjusting a time window of the time window scheduling framework to obtain a dynamically adjusted time window; inputting the state space and the action space into an improved double-delay depth deterministic strategy gradient algorithm for training to obtain a strategy generator; and performing multimodal transport scheduling decision generation based on the demand predictor, the dynamically adjusted time window and the strategy generator to obtain a multimodal transport scheduling scheme, so that efficient and flexible scheduling decision generation is realized, and the operation efficiency of multimodal transport is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent scheduling, and in particular to a multimodal transport intelligent scheduling optimization method, device, equipment and storage medium. Background Art

[0002] As an efficient and environmentally friendly mode of logistics and transportation, traditional multimodal transport scheduling methods can no longer meet the current rapidly changing market needs. The complexity of the multimodal transport network, the diversity of transportation modes and the uncertainty of cargo demand have brought huge challenges to scheduling optimization.

[0003] Existing multimodal transport scheduling methods often have several major problems: first, it is difficult to effectively handle large-scale, dynamically changing transportation networks, resulting in low scheduling efficiency; second, insufficient consideration is given to the coordination and balance between multiple modes of transportation, resulting in low resource utilization; third, the lack of accurate prediction of future freight demand makes scheduling decisions lack adaptability and foresight. Summary of the invention

[0004] The present invention provides a multimodal transport intelligent scheduling optimization method, device, equipment and storage medium, which realizes efficient and flexible scheduling decision generation and improves the operational efficiency of multimodal transport.

[0005] In a first aspect, the present invention provides a multimodal transport intelligent scheduling optimization method, the multimodal transport intelligent scheduling optimization method comprising: Model multiple transportation nodes and transportation paths to obtain the network topology; Divide the scheduling cycle into time periods based on the network topology to obtain a time window scheduling framework; The historical freight data is input into a multi-layer feed-forward neural network for training to obtain a demand forecaster; Dynamically adjusting the time window of the time window scheduling framework to obtain a dynamically adjusted time window; The state space and action space are input into the improved double-delayed deep deterministic policy gradient algorithm for training to obtain a policy generator; Based on the demand forecaster, the dynamically adjusted time window and the strategy generator, a multimodal transport scheduling decision is generated to obtain a multimodal transport scheduling solution.

[0006] In a second aspect, the present invention provides a multimodal transport intelligent scheduling optimization device, the multimodal transport intelligent scheduling optimization device comprising: A modeling module is used to model multiple transportation nodes and transportation paths to obtain a network topology structure; A division module, used to divide the scheduling period into time periods based on the network topology structure to obtain a time window scheduling framework; A training module, used to input historical freight data into a multi-layer feedforward neural network for training to obtain a demand forecaster; An adjustment module, used for dynamically adjusting the time window of the time window scheduling framework to obtain a dynamically adjusted time window; A processing module, used for inputting the state space and the action space into the improved double-delayed deep deterministic policy gradient algorithm for training to obtain a policy generator; A decision module is used to generate a multimodal transport scheduling decision based on the demand forecaster, the dynamically adjusted time window and the strategy generator to obtain a multimodal transport scheduling solution.

[0007] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned multimodal transport intelligent scheduling optimization method.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, wherein instructions are stored in the computer-readable storage medium, and when the computer-readable storage medium is run on a computer, the computer executes the above-mentioned multimodal transport intelligent scheduling optimization method.

[0009] In the technical solution provided by the present invention, a complete network topology is obtained by comprehensively modeling multiple transport nodes and transport paths. The scheduling cycle is divided into time periods based on the network topology, and a dynamic adjustment mechanism is introduced, so that the scheduling scheme can better adapt to changes in actual transportation demand. A multi-layer feedforward neural network is used to train historical freight data, and an adaptive adjustment mechanism is introduced to improve the accuracy of prediction of future freight demand. The policy generator obtained by training with an improved double-delay deep deterministic policy gradient algorithm can generate high-quality scheduling decisions in a complex state space. By introducing multi-mode coordination factors and effect evaluation indicators, while optimizing the total transportation cost and time, the balance between different modes of transportation is also considered. Mechanisms such as dynamic time window adjustment and adaptive learning rate are adopted to enable the scheduling system to better adapt to environmental changes and uncertainties. Based on the demand predictor, the dynamically adjusted time window and the policy generator, the multimodal transport scheduling scheme can be generated and updated in real time, which improves the response speed of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0011] Figure 1 A schematic diagram of the steps of the multimodal transport intelligent scheduling optimization method in an embodiment of the present invention; Figure 2 Schematic diagram of the structure of the multimodal transport intelligent scheduling optimization device in an embodiment of the present invention. DETAILED DESCRIPTION

[0012] Embodiments of the present invention provide a method, apparatus, device and storage medium for intelligent scheduling optimization of multimodal transport. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 An embodiment of the multimodal transport intelligent scheduling optimization method in the embodiment of the present invention includes: Step S1, modeling multiple transport nodes and transport paths to obtain a network topology structure; It is understandable that the execution subject of the present invention may be a multimodal transport intelligent scheduling optimization device, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0014] Specifically, the attributes of each transport node are defined to form a comprehensive set of node attributes. The node attribute set covers multiple key elements of each node, such as location coordinates, node type, and processing capacity indicators. The location coordinates are used to determine the specific location of the node in the geographic space, while the node type distinguishes the functional role of the node in the multimodal transport network, such as whether it is a port, railway station, or distribution center. The processing capacity indicator reflects the upper limit of the cargo flow that the node can handle within a certain period of time, such as the speed of loading and unloading cargo, storage capacity, etc. At the same time, the attributes of each transport path are defined to obtain a complete set of path attributes. The path attribute set includes the starting point and end point of the transport path, that is, the two nodes connected by the path. This information helps determine the specific route for the transportation of goods from one node to another. The mode of transportation describes the type of transportation used by the path. The transportation time refers to the time required to transport goods on the path, and the transportation cost reflects the economic cost of transporting goods on the path, including various related expenses such as fuel costs and tolls. The upper capacity value limits the maximum cargo flow that the path can carry per unit time. These attributes together describe the characteristics of the transport path. Based on the attribute information of nodes and paths, as well as actual geographic information and transportation network data, the connection relationship between each transportation node is determined to form a complete multimodal transport network topology. The historical freight data is collected and sorted to obtain the cargo inflow and outflow of each node and the actual transportation volume of each transportation path. The accumulation and analysis of historical data help identify the high-frequency transportation paths and nodes prone to bottlenecks in the network. The objective function is constructed based on the total transportation cost and total transportation time to obtain the network optimization target. The objective function is expressed as: F=min(α total transportation cost + β total transportation time), where α and β are the weight coefficients of transportation cost and transportation time, respectively. By minimizing the objective function, under certain weight settings, the optimal balance between overall transportation cost and time is sought to maximize the overall benefits of the network. The weight coefficients α and β in the objective function are initialized to obtain the initial weight value. The weight value is set based on the operational requirements and management objectives of the actual transportation network. For example, in time-sensitive transportation scenarios, a higher weight is given to the total transportation time, while in cost-sensitive scenarios, a higher weight is given to the total transportation cost. Based on the node attribute set and the path attribute set, various constraints of the network are constructed to obtain the initial network constraints. The constraints include the upper limit of the node's processing capacity, the capacity limit of the path, the priority requirements of a specific path, etc., to ensure that the solution in the optimization process is feasible and meets the physical and management limitations of actual operations. The network optimization goal, the initial weight value and the initial network constraints are combined to obtain a complete network topology.

[0015] Step S2: divide the scheduling cycle into time periods based on the network topology to obtain a time window scheduling framework; Specifically, based on the network topology, the entire scheduling cycle T is evenly divided into several discrete time periods {t1, t2, ..., tK}, where K is the number of preset time periods. The continuous scheduling cycle is discretized into several time segments that can specifically perform scheduling operations, which is convenient for subsequent scheduling decisions and optimization. The length of each time period is determined according to the actual transportation demand and scheduling flexibility, so that the scheduling task can be effectively processed within a time period without being too long or too short. For each transportation node Ni, a state vector is defined in each time period tk to obtain a node state vector Si,k. The definition of the node state vector includes two key elements: the current cargo volume Qi,k and the remaining processing capacity Ri,k. The current cargo volume Qi,k represents the amount of cargo existing on the node Ni in the time period tk, reflecting the cargo demand status or cargo flow status of the node in the time period. The remaining processing capacity Ri,k represents the processing margin of the node in the time period, that is, how much cargo the node can still process in the current time period. These two parameters together constitute the node state vector, which dynamically reflects the load situation and processing capacity of the node in each time period. For each transport path Rj, a state vector is defined in each time period tk to obtain the path state vector Pj,k. The path state vector mainly includes the current transport volume Vj,k and the remaining capacity Uj,k. The current transport volume Vj,k represents the amount of goods actually being transported on the path Rj in the time period tk, while the remaining capacity Uj,k represents the maximum cargo flow that the path can still carry in the current time period. These two indicators can reflect the usage of each path and their carrying capacity in real time, so that the existing capacity resources can be fully utilized when making scheduling decisions, while avoiding path overload or resource waste. The decision variable xi,j,k is introduced to represent the amount of goods sent from node Ni to node Nj in the time period tk, and the scheduling decision variable set is obtained. The scheduling decision variable set is a collective description of all possible scheduling operations in the entire scheduling system. By rationally configuring these variables, the optimal scheduling of the entire transportation network is achieved. These decision variables are subject to various constraints, including node capacity constraints, path capacity constraints, and cargo balance constraints. The node capacity constraint stipulates that the processing capacity of each node in each time period cannot exceed its upper limit, while the path capacity constraint limits the cargo transportation volume of each transportation path in each time period to not exceed its maximum capacity. The cargo balance constraint ensures that the inflow and outflow of cargo between each node and time period are balanced, that is, there cannot be an increase or decrease in cargo out of thin air. The constraints together constitute the time window scheduling constraint set to ensure the feasibility of the scheduling scheme in all aspects. In order to meet the timeliness requirements in actual transportation, the earliest shipment time TEi and the latest arrival time TLi are set for each batch of cargo, and the time window constraint TEi ≤ tk ≤ TLi - tj is obtained, where tj is the transportation time from node i to node j.The time window constraint ensures that each batch of goods can complete the transportation task within the specified time range, avoiding the early or delayed arrival of goods due to unreasonable scheduling arrangements, thereby affecting the overall logistics efficiency. The optimization objective function is refined and expressed as F=min(∑i,j,k(cjxi,j,k) +γ∑imax(0, Qi,K - Qi,0)), where cj is the unit cost of transportation from node Ni to node Nj, and γ is the penalty coefficient for goods that are not transported on time. The design of this objective function considers the trade-off between minimizing transportation costs and the penalty for not completing transportation tasks. By minimizing the transportation cost item, the transportation cost is minimized while ensuring timeliness. The penalty item for not completing transportation on time is to penalize the amount of goods that cannot be transported in the last time period, encouraging the scheduling strategy to complete the transportation task of goods within the specified time as much as possible. The discrete time period, node state vector, path state vector, scheduling decision variable set, time window scheduling constraint set, time window constraint and optimization objective function are combined to form a complete time window scheduling framework.

[0016] Step S3, inputting historical freight data into a multi-layer feedforward neural network for training to obtain a demand forecaster; Specifically, the historical freight data is preprocessed to convert the raw data into standardized input features that can be effectively learned and understood by the neural network model. The standardized input features include historical freight volume, seasonal factors, economic indicators, and weather conditions, which comprehensively reflect the multi-dimensional influencing factors of freight demand. In the preprocessing process, the raw data is cleaned, outliers and missing values ​​are removed, and all numerical features are standardized to make their mean 0 and variance 1, eliminating the impact of different feature scale differences on model training. A multi-layer feedforward neural network structure is designed to form a neural network model. The neural network model consists of an input layer, multiple hidden layers, and an output layer, in which the number of neurons in the input layer matches the dimension of the input feature, and each neuron corresponds to an input feature. The hidden layer part adopts a multi-layer structure to capture the complex nonlinear relationship between the input features. The number of neurons in each hidden layer is set according to the complexity of the actual problem. In order to enhance the expressive power of the model, the hidden layer uses the ReLU activation function. The nonlinear characteristics of the ReLU function enable the network to learn more complex patterns and also alleviate the gradient disappearance problem. The output layer uses a linear activation function because demand forecasting is a regression problem. The linear activation function can directly output the predicted value, which better reflects the continuity and changing trend of freight demand. After the model structure design is completed, the standardized input features are divided into a training set and a validation set in a ratio of 8:2, which effectively prevents the model from overfitting during the training process and facilitates the evaluation of the model's performance on unseen data. The training set data is normalized and all feature values ​​are scaled to the interval [0,1] to make the gradient smoother during the model training process, which helps to improve the training efficiency and model stability. The normalized training data is input into the neural network model and trained using the back propagation algorithm and Adam optimizer. The back propagation algorithm guides the update of the model parameters by calculating the gradient of the loss function relative to the model parameters, thereby gradually reducing the prediction error. The Adam optimizer is an adaptive learning rate optimization algorithm that can adaptively adjust the learning rate by calculating the first-order and second-order moment estimates and quickly converge to the global optimal solution. The mean square error is used as the loss function during training. This loss function measures the average square difference between the model prediction value and the true value, reflecting the accuracy of the model prediction. In order to prevent the model from overfitting during training, the dropout technique is applied to the model parameters. In each training iteration, part of the output of the hidden layer neurons is randomly discarded, and the dropout rate is set to 0.5, that is, 50% of the neuron outputs will be temporarily ignored in each training to prevent the model from over-relying on the weights of certain specific neurons. After the training is completed, the k-fold cross-validation method is used to comprehensively evaluate the model, and the k value is set to 5.Cross-validation divides the data set into k non-overlapping subsets, uses one of the subsets as the validation set, and uses the remaining k-1 subsets as the training set for training and validation, effectively reducing the random error caused by the data division method and obtaining more stable model performance indicators. The performance indicators include prediction accuracy, mean absolute error, and root mean square error, which measure the prediction performance of the model in different aspects. In order to cope with the dynamic changes in freight demand and the performance fluctuations of the model in practical applications, an adaptive adjustment mechanism is designed to improve the prediction effect of the model. The mechanism includes triggering model retraining when the prediction error of the model exceeds the preset threshold, and using an adaptive learning rate to update the weights of the neural network. The adaptive learning rate is dynamically adjusted using the formula η(t) =η0 / (1+λt), where η0 is the initial learning rate, λ is the decay coefficient, t is the number of training steps, and η(t) is the adaptive learning rate. As the number of training steps t increases, the learning rate gradually decays, so that the model can converge to the local optimal solution more stably in the later stage of training. The adaptive adjustment mechanism enables the demand forecaster to maintain high forecasting performance in the face of different freight environment changes, and ultimately obtains a demand forecaster model with stable performance and capable of adapting to actual demand fluctuations.

[0017] Step S4, dynamically adjusting the time window of the time window scheduling framework to obtain a dynamically adjusted time window; Specifically, a time window adjustment factor δi,k is defined, which can quantify the adjustment demand in each time period. The adjustment amount calculation formula is δi,k=min(ρ*(Di,k-Ci,k),δmax), where Di,k represents the predicted demand in time period tk, Ci,k represents the current transportation or processing capacity, ρ is the adjustment coefficient, which is used to control the sensitivity of the adjustment amount, and δmax is the maximum adjustment amount, which is used to limit the amplitude of a single adjustment and prevent excessive adjustment from causing scheduling chaos. This formula considers the gap between actual demand and existing capacity and quantifies it into an operational adjustment value, providing basic data for the adjustment of subsequent time windows. After the adjustment amount δi,k is calculated, the time window is updated based on the adjustment amount to obtain the updated time window. The update formula is TEi,k+1=TEi,k-max(0,δi,k) and TLi,k+1=TLi,k+max(0,δi,k), where TEi,k and TLi,k represent the earliest shipment time and the latest arrival time in time period tk, respectively. By adjusting the earliest shipment time and the latest arrival time, the duration and scheduling flexibility of each time window can be flexibly controlled to adapt to the dynamically changing transportation demand and resource conditions. For example, when the predicted demand is higher than the current capacity, δi,k is a positive value, the earliest shipment time TEi,k is reduced accordingly, and the latest arrival time TLi,k is increased accordingly, so as to free up more time to handle additional demand; when the predicted demand is lower than the current capacity, δi,k is a negative value, and the time window will be shortened accordingly to reduce resource waste. Constraints are set for the updated time window to avoid excessive adjustment of the time window beyond a reasonable range. The specific constraints are TEmin≤TEi,k≤TEmax and TLmin≤TLi,k≤TLmax, where TEmin and TEmax are the minimum and maximum limits of the earliest shipment time, and TLmin and TLmax are the minimum and maximum limits of the latest arrival time, respectively. Upper and lower limit constraints ensure that the dynamic adjustment of the time window is carried out within a controllable range, preventing the time window from being too long or too short, resulting in a decrease in scheduling efficiency or a decrease in service level. At the same time, in order to achieve better coordination and resource balance between different transportation modes, a multimodal coordination factor μj is introduced. The calculation formula of the multimodal coordination factor is μj=(cj / cmax)*(tj / tmax), where cj is the transportation cost of path j, tj is the transportation time of path j, cmax and tmax are the maximum cost and maximum time of all paths respectively. By standardizing and combining the two factors of transportation cost and time, μj reflects the comprehensive balance between different transportation paths, and can effectively regulate the resource allocation and use of different modes in scheduling decisions. The multimodal coordination factor μj is integrated into the optimization objective function to obtain a new optimization objective function F'.The new optimization objective function is expressed as F'=min(∑i,j,k(μjcjxi,j,k)+γ*∑imax(0,Qi,K-Qi,0)), which adapts to the coordinated scheduling needs of different modes of transport in the multimodal transport network and improves the overall scheduling efficiency of the system. In order to evaluate the effect of the dynamically adjusted time window, the effect evaluation index E is designed, and the evaluation formula E=w1*(actual cost / expected cost)+w2*(actual time / expected time) is given, where w1 and w2 are weight coefficients, representing the importance of cost and time in the evaluation index respectively. The effect of time window adjustment is quantified by calculating the ratio of actual cost to expected cost and actual time to expected time. When E>1, it indicates that the actual cost and time exceed expectations and need to be further optimized; when E<1, it indicates that the actual situation is better than expected. Based on the calculation results of the evaluation formula, the time window adjustment strategy for the next cycle is adaptively adjusted to more accurately match the changes in transportation demand and resource supply. When E>1, it indicates that the actual cost and time are both higher than expected. At this time, the time window adjustment of the next cycle should be increased to better adapt to demand changes and alleviate resource pressure; when E<1, it means that the current scheduling strategy is ideal, and the adjustment range should be appropriately reduced to maintain stable operation of the system. The adaptive adjustment strategy can flexibly adjust the time window according to the actual operation situation and improve the dynamic response capability of the scheduling strategy. The constrained time window, the new optimization objective function and the adaptive adjustment strategy are combined to obtain the dynamically adjusted time window. This time window can dynamically adapt to changes in transportation demand in different time periods and realize efficient allocation of resources between different transportation modes. Through continuous optimization through the adaptive adjustment strategy, the overall scheduling level and operational efficiency of the multimodal transport system can be improved.

[0018] Step S5, inputting the state space and the action space into the improved double-delayed deep deterministic policy gradient algorithm for training to obtain a policy generator; Specifically, the state space S and the action space A are defined. The state space S includes the current cargo volume of all transportation nodes, the current transportation volume of all transportation paths, and the time window information, reflecting the current state of the entire multimodal transport network. The action space A represents the transportation volume allocation decision for each path, that is, the decision to transport a specific amount of cargo from a node to another node through a path in each time period. The setting of the action space determines the behavior choice of the scheduling system under different states. A deep neural network is constructed to approximate the policy and value functions. In the improved DDPG algorithm, an Actor network fθ(s) and two Critic networks Qφ1(s,a) and Qφ2(s,a) are constructed respectively. The network structure is used to estimate the policy value and value of taking action a under a given state s. The Actor network fθ(s) accepts the current state s as input and outputs a specific action a, that is, the policy π(s), while the Critic networks Qφ1 and Qφ2 are used to estimate the Q value of a given state-action pair (s,a). Through the design of this structure, the network continuously adjusts the strategy π(s) and the value function Q(s,a) during the training process to achieve the learning of the optimal scheduling strategy. In order to improve the stability and convergence speed of network training, the target network fθ' and Qφ1', Qφ2' are introduced, and the parameters of the target network are updated through the soft update mechanism. The soft update mechanism adopts the parameter update formula θ'←τθ+(1-τ)θ' and φi'←τφi+(1-τ)φi', where τ is the soft update coefficient. The introduction of the soft update mechanism avoids the training instability caused by the rapid change of the target network parameters, and ensures the smooth convergence of the network during the training process. In the action generation process, in order to explore more state-action space in the early stage of training, Gaussian noise ε is introduced and added to the strategy generation to obtain the exploration strategy μ(s). The specific calculation formula is μ(s)=π(s)+ε, where ε obeys a normal distribution with a mean of 0 and a standard deviation of σ. By adding noise when generating actions, the strategy generator explores the state space as much as possible in the early stage of training, thereby avoiding falling into the local optimal solution. An adaptive noise attenuation mechanism is designed to control the balance between exploration and exploitation during training. The noise standard deviation update formula is σt=σ0*exp(-λt), where σ0 is the initial noise standard deviation, λ is the decay rate, and t is the number of training steps. As training progresses, the noise standard deviation gradually decreases, and the policy generator gradually shifts from extensive exploration in the early stage to precise policy optimization in the later stage. In order to make full use of the existing experience data, an experience replay buffer D is defined, which is used to store the state-action-reward-next state quadruplets accumulated during training. During training, a sampling mechanism is used to randomly extract mini-batch data from D for training, which breaks the temporal correlation between data and improves the efficiency and stability of training.The randomly extracted mini-batch data includes multiple state-action pairs and their corresponding rewards and next state information. Through batch sampling training, the network parameters are updated more comprehensively and the generalization ability of the model is improved. In the update of the Critic network, the calculation formula of the target Q value is y=r+ω*min(Qφ1'(s',π(s')),Qφ2'(s',π(s'))), where r is the current reward, s' is the next state after executing action a, and ω represents the discount factor. The parameter φi of the Critic network is updated by minimizing the difference between the target Q value and the current Q value. The specific update formula is φi←φi-η∇φiL(φi), where L(φi) is the loss function. At the same time, the update formula of the Actor network is θ←θ+η∇aQφ1(s,a)|a=π(s)∇θπ(s), that is, the parameter θ of the policy network is updated by maximizing the Q value to improve the quality of the actions generated by the policy network. In order to ensure that the learning rate can be adaptively adjusted during training, an adaptive learning rate adjustment mechanism is designed. The learning rate update formula is η=η0 / (1+β*t), where η0 is the initial learning rate, β is the decay coefficient, and t is the number of training steps. As the number of training steps increases, the learning rate gradually decreases, and the pace of network parameter update becomes more gentle, making the training process more stable and avoiding oscillation. The neural network structure, soft update mechanism, exploration strategy, noise standard deviation update formula, experience replay sampling mechanism, network update formula and learning rate update formula are organically combined to form a complete strategy generator. Based on the input state space and action space, the strategy generator can continuously adjust and optimize the multimodal transport scheduling strategy through the training process of deep reinforcement learning, and realize intelligent scheduling decisions for multiple transportation modes in complex transportation networks.

[0019] Step S6: Based on the demand forecaster, the dynamically adjusted time window and the strategy generator, a multimodal transport scheduling decision is generated to obtain a multimodal transport scheduling solution.

[0020] Specifically, the key parameters of the system are initialized, including the basic structure of the multimodal transport network, the initial time window setting, and the initial weights of the neural network used for strategy generation. The initialization of the network structure includes determining the number of transport nodes in the multimodal transport system and the path connection relationship between them. These connection relationships reflect the connection mode and path selection between different transport modes. The initialization of the time window includes the initial earliest shipment time (TE) and the latest arrival time (TL) of each node. The time window provides a time boundary for subsequent scheduling decisions. The initialization of the neural network weights includes the initial weight values ​​of the actor network and the two critic networks. These weights affect the exploration behavior of the strategy generator in the early stage of training. The current system state is input into the demand forecaster to predict the freight demand and capacity supply in the next cycle. The current state includes the current cargo volume of each node, the current transportation volume of each transportation path, and the information of the current time window. This information can reflect the real-time operation status of the system. Based on historical data and current state information, the demand forecaster predicts the freight demand and capacity supply values ​​in the next cycle through the operation of a multi-layer feedforward neural network. The prediction results include the inflow and outflow of goods for each node in the future time period, as well as the future capacity supply of each path. Based on the prediction results of the demand predictor, the dynamically adjusted time window is updated to obtain a more accurate time window setting. The update process uses the time window adjustment formula proposed in claim 5, that is, the earliest shipment time TE and the latest arrival time TL are adaptively adjusted by adjusting the factors δi,k. When the predicted demand for a node in the future time period exceeds the current capacity, the earliest shipment time of the node will be reduced or the latest arrival time will be increased to provide a time window for more transportation needs, and vice versa. Through dynamic adjustment, the time window can respond more flexibly to fluctuations in transportation demand and improve the scheduling efficiency and service level of the entire system. After the time window update is completed, the current state st is input into the policy generator to generate the scheduling decision at under the current state. The policy generator learns and generates policies based on the improved double-delayed deep deterministic policy gradient algorithm (DDPG). The specific calculation process is at=μ(st)=π(st)+ε, where π(st) is the scheduling policy generated by the Actor network according to the current state st, and ε is the exploration noise, which obeys the normal distribution with a mean of 0 and a standard deviation of σ. By introducing exploration noise, the exploration ability of the policy generator during the training process is enhanced to avoid premature convergence to the local optimal solution. The generated scheduling decision at includes the transportation volume allocation of each transportation path in each time period, that is, how much cargo is transported from one node to another in different time periods. Based on the scheduling decision, the transportation process of the entire multimodal transport system is simulated. The simulation process includes updating the cargo volume of each node and the transportation volume of each path, and calculating the total transportation cost and total transportation time.In each time period, according to the scheduling decision, the cargo transportation volume on each path is updated to the corresponding node, and the transportation cost and time overhead caused by the scheduling scheme are recorded. This simulation process can help verify the actual effect of the current scheduling strategy and provide reference data for subsequent strategy optimization. After the transportation simulation is completed, the reward value rt of the scheduling strategy is calculated, and the expression of the reward function is obtained. The reward function is designed as rt=-F'(st,at), where F' is the new optimization objective function. By minimizing the total transportation cost and total transportation time and introducing considerations of resource utilization and service level, the reward function can effectively reflect the pros and cons of the scheduling strategy. The introduction of the negative sign indicates that the system hopes to minimize the cost and time, so that the reward value is as large as possible, so as to guide the strategy generator in a direction that is conducive to system performance optimization. After obtaining the reward value of the current decision, observe the new state st+1 after executing the current scheduling decision, update the cargo volume of each node, the transportation volume of each path, and the time window information, so as to form a complete state transition sample (st,at,rt,st+1). The state transition sample can completely record the whole process from the current state to the new state by executing a certain scheduling decision, including the gains or losses brought by the scheduling decision. The state transition samples are stored in the experience replay buffer D. The size of the experience pool is the preset value M. When the experience pool is full, the old samples will be replaced by new samples to ensure the diversity of samples. Through this mechanism, the time correlation problem of samples can be effectively avoided and the learning effect of the policy generator can be improved. During the training process, N mini-batch data are randomly sampled from the updated experience pool to obtain a training sample set. The randomly sampled N sample sets include several state transition samples (st, at, rt, st+1). These samples can provide rich state-action pairs and their corresponding feedback information, thereby supporting the training process of the policy generator. Based on the training sample set, the network parameters of the policy generator are updated. The specific update process includes first calculating the target Q value. The calculation formula of the target Q value is y=r+γ*min(Qφ1'(s',π(s')),Qφ2'(s',π(s'))), where r is the current reward, γ is the discount factor, Qφ1' and Qφ2' are the target Critic networks, and π(s') is the target Actor network. By minimizing the difference between the target Q value and the actual Q value, the Critic network parameters φi are updated. The specific update formula is φi←φi-η∇φiL(φi), where L(φi) is the loss function and η is the learning rate. The Actor network parameters are updated every d steps. The update formula is θ←θ+η∇aQφ1(s,a)|a=π(s)∇θπ(s), that is, the parameters θ of the policy network are updated by maximizing the Q value, thereby improving the action quality generated by the policy network. The update results of the policy generator are applied to the entire multimodal transport system to generate a specific scheduling plan.In the process of generating a scheduling plan, the cargo volume of each node and the transportation capacity of each path are taken into consideration, as well as the dynamic adjustment of the time window and the prediction of future demand, so that the scheduling plan can be more in line with the actual situation.

[0021] In the embodiment of the present invention, a complete network topology is obtained by comprehensively modeling multiple transport nodes and transport paths. The scheduling cycle is divided into time periods based on the network topology, and a dynamic adjustment mechanism is introduced so that the scheduling scheme can better adapt to changes in actual transportation demand. A multi-layer feedforward neural network is used to train historical freight data, and an adaptive adjustment mechanism is introduced to improve the accuracy of prediction of future freight demand. The policy generator obtained by training with an improved double-delayed deep deterministic policy gradient algorithm can generate high-quality scheduling decisions in a complex state space. By introducing multi-mode coordination factors and effect evaluation indicators, while optimizing the total transportation cost and time, the balance between different modes of transportation is also considered. Mechanisms such as dynamic time window adjustment and adaptive learning rate are adopted to enable the scheduling system to better adapt to environmental changes and uncertainties. Based on the demand predictor, the dynamically adjusted time window and the policy generator, the multimodal transport scheduling scheme can be generated and updated in real time, which improves the response speed of the system.

[0022] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Define the attributes of each transport node to obtain a node attribute set, which includes location coordinates, node type and processing capacity indicators; Define the attributes of each transportation route to obtain a set of route attributes, which includes the starting point, end point, transportation method, transportation time, transportation cost and transportation capacity upper limit; Based on the actual geographic information and transportation network data, the connection relationship between each node is determined to obtain a complete multimodal transport network topology. The historical freight data is collected and sorted to obtain the cargo inflow and outflow of each node and the actual transportation volume of each transportation route. Based on the total transportation cost and total transportation time, the objective function is constructed to obtain the network optimization target. The objective function is F=min(α*total transportation cost+β*total transportation time); Initialize the weight coefficients α and β in the objective function to obtain the initial weight value, and build network constraints based on the node attribute set and the path attribute set to obtain the initial network constraints; The network optimization objective, initial weight value and initial network constraint are combined to obtain the network topology.

[0023] Specifically, the attributes of each transport node are defined to form a node attribute set. The construction of the node attribute set usually includes three core elements: location coordinates, node type, and processing capacity index. The location coordinates are used to describe the specific location of each node in the geographic space. They are expressed in latitude and longitude coordinates and can accurately reflect the distribution of nodes in the actual geographical environment. The node type is used to distinguish the functional roles of different nodes, such as ports, railway hubs, distribution centers, etc. Different types of nodes have different functions and roles, and participate in different transportation modes. The processing capacity index represents the amount of cargo that each node can handle per unit time, reflecting the loading, unloading, storage and transshipment capabilities of the node. The attributes of each transport path are defined to construct a path attribute set. The path attribute set covers a number of key data in the transportation process, including the starting point, end point, transportation mode, transportation time, transportation cost and transportation capacity limit. The connection relationship between each node is determined based on the actual geographic information and transportation network data to obtain a complete multimodal transport network topology. Integrate multiple data sources, such as geographic information system (GIS) data, traffic management data, and historical transportation data, to accurately describe the physical connection between nodes. For each node, the number of connected paths and their types directly affect the flexibility and efficiency of transportation. By integrating this information, a comprehensive network topology is formed to reflect the actual transportation network layout and function. Historical freight data is collected and sorted to obtain the cargo inflow and outflow of each node, as well as the actual transportation volume of each transportation path. Historical data helps to identify frequently used paths and nodes, predict future transportation demand, and provide real operational data for network optimization. In order to optimize the entire transportation network, an objective function is constructed to measure the pros and cons of different options. The objective function should take into account the total transportation cost and the total transportation time, and the goal is to minimize the weighted sum of these two factors. Define the objective function as: ; in, represents the total transportation cost, Indicates the total transportation time, and is the corresponding weight coefficient. Total transportation cost Includes the cost of transporting goods on all routes, calculated by multiplying the transportation cost of each route by the transportation volume. For example, for a route , if the transportation cost is , the amount of cargo transported on the path is , then the total cost of the path is Total shipping time is the total time required to transport goods on all routes. The transportation time multiplied by the transportation volume can reflect the time burden of the route. The objective function optimizes the entire system by adjusting the balance between transportation cost and time. In order to enhance the flexibility of the objective function, the weight coefficient and Perform initialization settings. For example, in some time-sensitive transportation tasks, setting Larger, Small, thus giving priority to shortening transportation time; on the contrary, in cost-sensitive scenarios, increasing The weights of the paths can be adjusted to reduce the total transportation cost. By setting reasonable weights, the system can balance the needs of different scenarios and improve the flexibility and accuracy of optimization. Based on the node attribute set and the path attribute set, the corresponding network constraints are set. The initial network constraints usually include the processing capacity limit of the node and the upper limit of the path capacity. For example, if the daily processing capacity of a node is 5,000 tons, the total inflow and outflow of goods at the node cannot exceed this upper limit. Similarly, the upper limit of the path capacity cannot be exceeded, otherwise it will cause transportation delays or system crashes. Constraints ensure the feasibility of the optimization scheme in actual operation and avoid unreasonable scheduling arrangements. Other constraints include the priority of specific paths, time window restrictions, etc. The constructed network optimization objectives, initialized weight values, and initial network constraints are combined to obtain a complete network topology structure, which reflects the nodes, paths and their attributes in the multimodal transport system, and provides information for the optimal scheduling of the system through the setting of objective functions and constraints.

[0024] In a specific embodiment, the process of executing step S2 may specifically include the following steps: The entire scheduling period T is evenly divided into K discrete time periods {t1, t2, ..., tK}, where K is the number of preset time periods; Define the state vector for each transport node Ni in each time period tk, and obtain the node state vector Si,k, which includes the current cargo volume Qi,k and the remaining processing capacity Ri,k; Define the state vector for each transport route Rj in each time period tk to obtain the route state vector Pj,k, which includes the current transport volume Vj,k and the remaining transport capacity Uj,k; Introduce decision variables xi,j,k to represent the amount of cargo shipped from node Ni to node Nj in time period tk, and obtain the set of scheduling decision variables; Based on the node capacity constraints, path capacity constraints and cargo balance constraints, the constraints are constructed to obtain the time window scheduling constraint set; For each batch of goods, the earliest shipping time TEi and the latest arrival time TLi are set, and the time window constraint TEi≤tk≤TLi-tj is obtained, where tj is the transportation time from node i to node j; The objective function is refined into F=min(∑i,j,k(cj*xi,j,k)+γ*∑imax(0,Qi,K-Qi,0)) to obtain the optimized objective function, where γ is the penalty coefficient for goods that are not transported on time; The discrete time period, node state vector, path state vector, scheduling decision variable set, time window scheduling constraint set, time window constraint and optimization objective function are combined to obtain the time window scheduling framework.

[0025] Specifically, the entire scheduling cycle Divide evenly, and get Discrete time periods ,in is the number of preset time periods. The continuous scheduling cycle is divided into several discrete time slices, so that each time period can make independent scheduling decisions. Discretization processing makes it easy to transform complex continuous scheduling problems into several easy-to-solve discrete optimization problems. In each time period Define the state vector Get the node state vector. Node state vector Including current cargo volume and remaining processing capacity Current cargo volume refers to the time period node The total amount of cargo on the node reflects the cargo storage status of the node. It indicates the remaining cargo volume that the node can handle during this time period, that is, the cargo load that the node can still bear in transportation, loading and unloading operations. In time period The current cargo volume of the node is 500 tons, and its maximum handling capacity is 1000 tons / hour. If there are no other influencing factors, the node The remaining processing capacity is 500 tons. Through the definition of state vector, the cargo status and processing capacity of each node are described in different time periods, providing accurate data support for scheduling decisions. Similarly, for each transportation path In each time period Define the state vector Get the path state vector. Path state vector Including current transport volume and remaining capacity Current transport volume refers to the time period path The actual amount of cargo transported on the route reflects the actual load of the route. It refers to the maximum amount of cargo that can still be transported by the route within the current time period. For example, if a railway transport route The upper limit of the transport capacity is 1000 tons / hour, and the current transport volume is 800 tons. By defining the path state vector, the usage of each transportation path is dynamically monitored to ensure that there is no overload or waste of resources during transportation. Introducing decision variables Indicates the time period Slave Node Send to Node The amount of cargo. The set of decision variables It includes the transportation decisions between all node pairs and is the core variable of the scheduling optimization problem. By optimizing these variables, the distribution and transportation volume of goods between nodes in each time period are determined to achieve the optimal scheduling of the entire multimodal transport system. In order to make the scheduling plan meet various constraints in actual operation, constraints are constructed based on node capacity constraints, path capacity constraints and cargo balance constraints to obtain the time window scheduling constraint set. The node capacity constraint stipulates that the cargo inflow and outflow of each node in each time period cannot exceed its processing capacity, that is, , which represents the node The total amount of goods inflow minus the total amount of goods outflow cannot exceed its remaining processing capacity. The path capacity constraint requires that the transportation volume on each transportation path cannot exceed its capacity limit, that is, , which represents the node To Node The cargo volume cannot exceed the path The cargo balance constraint requires that in each time period, the total inflow and outflow of cargo at all nodes should be balanced, that is, , which represents the node In time period The total outflow of goods should be equal to its current cargo volume. Constraints ensure that the scheduling plan is feasible in actual execution and complies with the physical and management constraints of the system. On the basis of satisfying the above constraints, the earliest shipping time is set for each batch of goods. and the latest arrival time , get the time window constraint ,in For slave nodes To Node The time window constraint ensures that each batch of goods can complete the transportation task within the specified time range, avoiding the situation where the goods arrive early or late due to unreasonable scheduling arrangements, thereby improving the logistics service level. The optimization objective function is refined to minimize the transportation cost and time during the scheduling process. The optimization objective function is expressed as: ; in, Indicates the path The transportation cost from node To Node The unit transportation cost, is a decision variable, representing the time period Slave Node Send to Node volume of cargo. is the penalty factor for goods that are not transported on time, reflecting the cost of delayed or uncompleted tasks. and Respectively represent nodes The amount of cargo in the last time period and the initial time period. The first part of the objective function reflects the minimization of total transportation costs, the second part It represents the penalty for failing to complete the transportation task in time, forcing the system to complete the cargo transportation task on time as much as possible. By solving the optimization objective function, the optimal cargo transportation scheduling plan is obtained under the premise of meeting all constraints. The discrete time period, node state vector, path state vector, scheduling decision variable set, time window scheduling constraint set, time window constraint and optimization objective function are organically combined to form a complete time window scheduling framework. This framework can dynamically adjust the scheduling plan of each node and path according to the actual operation of the multimodal transport system to ensure that the cargo can complete the transportation task within the optimal time and cost range.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: Preprocess the historical freight data to obtain standardized input features, including historical freight volumes, seasonal factors, economic indicators, and weather conditions; Design a multi-layer feedforward neural network structure to obtain a neural network model, which includes an input layer, multiple hidden layers and an output layer, where the hidden layer uses a ReLU activation function and the output layer uses a linear activation function; The standardized input features are divided into a training set and a validation set in a ratio of 8:2, and the training set data is normalized to obtain the normalized training data. The normalization process scales all feature values ​​to the [0,1] interval; The normalized training data is input into the neural network model, and the back propagation algorithm and Adam optimizer are used for training to obtain the initial model parameters. The mean square error is used as the loss function in the training process. Apply the dropout technique to the initial model parameters to obtain model parameters that prevent overfitting, and the dropout rate is set to 0.5; The k-fold cross-validation method was used to evaluate the model and obtain the model performance indicators. The k value was set to 5. The performance indicators included prediction accuracy, mean absolute error, and root mean square error. An adaptive adjustment mechanism is designed to obtain a demand forecaster. The adaptive adjustment mechanism includes: when the prediction error exceeds a preset threshold, the model retraining is triggered, and the neural network weights are updated using an adaptive learning rate, where the adaptive learning rate η(t)=η0 / (1+λt), η0 is the initial learning rate, λ is the attenuation coefficient, t is the number of training steps, and η(t) is the adaptive learning rate.

[0027] Specifically, the historical freight data is preprocessed to convert the raw data into standardized input features that can be understood and used by the neural network model. The input features include historical freight volume, seasonal factors, economic indicators and weather conditions. Historical freight volume reflects the freight transportation situation of each node and path in different time periods. Seasonal factors capture the fluctuation of transportation demand due to seasonal changes. Economic indicators include some macroeconomic data, such as GDP growth rate, industrial production index, etc., which reflect the impact of macroeconomic conditions on freight demand. Weather conditions include information such as temperature, rainfall, wind speed, etc. In the preprocessing process, these data are cleaned and outliers are processed, and missing values ​​are filled or deleted. All numerical features are standardized so that the mean of each feature is 0 and the variance is 1, so as to eliminate the impact of scale differences between different features on model training. After standardization, the input features are better utilized by the neural network model, improving the training efficiency and prediction accuracy of the model. A multi-layer feedforward neural network structure is designed to construct a neural network model. The model includes an input layer, multiple hidden layers and an output layer. The number of neurons in the input layer matches the dimension of the standardized input feature, and each neuron receives a feature value. The hidden layer is composed of several layers, each containing a certain number of neurons, which are used to capture the complex nonlinear relationship between input features. The activation function of the hidden layer uses the ReLU function, that is, ReLU , the activation function can effectively alleviate the gradient vanishing problem and improve the training speed and expression ability of the model. The output layer uses a linear activation function to output the predicted freight demand value. The form of the linear activation function is , suitable for regression problems, and can directly output predicted values ​​without additional nonlinear transformations. Through this network structure design, the model can learn complex mapping relationships under different feature combinations, providing accurate model support for demand forecasting. After completing the design of the neural network model, the standardized input features are divided into training sets and validation sets in a ratio of 8:2, that is, 80% of the data is used for model training and 20% of the data is used for model validation. The training set is used to learn model parameters, while the validation set is used to evaluate the model's generalization ability and performance on unseen data. In order to improve the training efficiency of the model, the training set data is normalized, that is, all feature values ​​are scaled to the [0,1] interval. The formula for normalization is: in, represents the original eigenvalue, and are the minimum and maximum values ​​of the feature, respectively. is the normalized eigenvalue. Normalization makes all input eigenvalues ​​in the same numerical range, which helps the model converge faster and avoids the adverse effects of eigenvalues ​​that are too large or too small on gradient calculations during model training. The normalized training data is input into the neural network model and trained using the back propagation algorithm and Adam optimizer. The back propagation algorithm gradually adjusts the model parameters to minimize the loss function by calculating the gradient of the loss function relative to the model parameters. The mean square error is used as the loss function, and its formula is: ; in, is the sample size, is the true value, is the predicted value of the model. The mean square error measures the average squared difference between the model's predicted value and the true value. The Adam optimizer is an adaptive learning rate optimization algorithm that adjusts the learning rate based on the first-order and second-order moment estimates of the gradient at each parameter update. The Adam optimizer calculates the gradient , that is, the loss function for the current model parameters The partial derivative of : ; in, Indicates that the loss function is at the current parameter The value at . The first-order moment (momentum) of the Adam optimizer on the gradient and the second moment (the square of the gradient) Perform exponentially weighted moving average, the formula is as follows: ; in, and are the decay rates of the first-order and second-order moment estimates, respectively, and are generally set to 0.9 and 0.999. and Perform bias correction to obtain the corrected value: ; ; The Adam optimizer uses the corrected gradient momentum and the squared corrected gradient Update model parameters ; in, is the learning rate, is a very small constant (usually ) is used to prevent zero division errors. The parameter update formula can maintain stable updates at different gradient scales to avoid training instability caused by too large or too small learning rates. After the model training is completed, the dropout technique is applied to the initial model parameters to prevent overfitting. Dropout is a regularization method that randomly discards some neurons. In each training iteration, the outputs of some neurons are randomly set to zero with a certain probability, reducing the model's dependence on certain specific features and improving the model's generalization ability. The dropout rate is set to 0.5, which means that in each training, 50% of the neurons will be randomly blocked, and only the other half of the neurons will participate in the training. In order to evaluate the performance of the model, the k-fold cross-validation method is used to comprehensively evaluate the model. The k-fold cross-validation divides the data set into k non-overlapping subsets, and uses one of the subsets as the validation set in turn, and the remaining k-1 subsets as the training set for training and validation. Repeat k times to effectively reduce the random error caused by data partitioning. Select k=5, that is, divide the data into 5 subsets, use one of the subsets as the validation set each time, and the other 4 subsets as the training set. Through cross-validation, more stable model performance indicators are obtained, including prediction accuracy, mean absolute error, and root mean square error. The formula for mean absolute error is: ; in, is the sample size, is the true value, For the predicted value, MAE measures the average absolute difference between the predicted value and the true value. The formula for the root mean square error is: ; RMSE is the square root of the mean square error, which can more intuitively reflect the average size of the prediction error. Through these two indicators, the prediction performance of the model in different aspects is comprehensively evaluated. In order to improve the dynamic adaptability of the model, an adaptive adjustment mechanism is designed. When the prediction error of the model exceeds the preset threshold, the model is retrained and the weights of the neural network are updated using an adaptive learning rate. The update formula of the adaptive learning rate is: ; in, is the initial learning rate, is the learning rate decay coefficient, is the number of training steps, is the adaptive learning rate. As the number of training steps increases As the learning rate increases, it gradually decays, so that the model can converge to the optimal solution more stably in the later stage of training. Through the adaptive adjustment mechanism, when the model prediction error is large, the learning rate and weight update strategy of the model are dynamically adjusted to continuously improve the prediction accuracy and stability of the model. A demand forecaster with stable performance and adaptability to changing environments is formed.

[0028] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Define the time window adjustment factor δi,k, and get the adjustment amount calculation formula, which is δi,k=min(ρ*(Di,k-Ci,k),δmax), where Di,k is the predicted demand, Ci,k is the current capacity, ρ is the adjustment coefficient, and δmax is the maximum adjustment amount; The time window is updated based on the adjustment calculation formula to obtain the updated time window. The update formula is TEi,k+1=TEi,k-max(0,δi,k), TLi,k+1=TLi,k+max(0,δi,k), where TEi,k is the earliest shipment time and TLi,k is the latest arrival time. Adjust the upper and lower limits of the updated time window to obtain the constrained time window. The constraints are TEmin≤TEi,k≤TEmax, TLmin≤TLi,k≤TLmax. Introducing the multimodal coordination factor μj, we get the multimodal balance formula: μj=(cj / cmax)*(tj / tmax), where cj is the transportation cost of path j, tj is the transportation time of path j, cmax and tmax are the maximum cost and maximum time of all paths respectively; Integrate the multi-mode coordination factor into the objective function to obtain a new optimization objective function F'. The new optimization objective function is F'=min(∑i,j,k(μj*cj*xi,j,k)+γ*∑imax(0,Qi,K-Qi,0)); Design effect evaluation index E, and obtain the evaluation formula, which is E=w1*(actual cost / expected cost)+w2*(actual time / expected time), where w1 and w2 are weight coefficients; Based on the calculation result of the evaluation formula, the time window adjustment strategy of the next cycle is adjusted to obtain an adaptive adjustment strategy. When E>1, the time window adjustment strength of the next cycle is increased, and when E<1, the adjustment amplitude of the next cycle is reduced. The constrained time window, the new optimization objective function and the adaptive adjustment strategy are combined to obtain a dynamically adjusted time window.

[0029] Specifically, define a time window adjustment factor , which is used to quantify the measure of time window adjustment at different nodes and time periods. The calculation formula of the time window adjustment factor is: ; in, Indicates the time period Internal Node The forecast demand represents how much cargo is expected to need to be handled or transported during that time period. Indicates the processing capacity of the current node, reflecting the maximum processing capacity of the node in this time period. Adjustment coefficient The sensitivity of the control adjustment determines how quickly the adjustment changes when there is a difference between the demand and the processing capacity. A higher setting will be more sensitive to differences in demand and capacity. is the maximum adjustment amount, which is used to limit the amplitude of each adjustment to prevent excessive adjustment from causing scheduling chaos. Through this formula, the time window adjustment amount required in each time period and each node is determined. Greater than current processing capacity When the adjustment If it is a positive value, it means that the time window needs to be increased; otherwise, it is zero or a negative value, which means that the time window does not need to be increased or even reduced. Based on the above adjustment amount calculation formula, the time window of each node is updated to obtain the updated time window. The update process is as follows: ; ; in, and Respectively represent nodes In time period The earliest shipment time and the latest arrival time. Update the time window, when adjusting the amount When it is positive, the earliest shipping time Decrease means that it is necessary to ship in advance to accommodate more transportation needs; the latest arrival time Increase means that the goods are allowed to arrive in a longer time to disperse the transportation pressure. If the adjustment amount is zero or negative, the time window remains unchanged or shortened. After updating the time window, set upper and lower limit constraints on the updated time window to avoid unreasonable scheduling caused by excessive adjustment of the time window. The constraints are: ; ; in, and are the minimum and maximum limits of the earliest shipping time, and The upper and lower limits ensure that the time window adjustment is carried out within a controllable range, preventing too long or too short time windows from affecting scheduling efficiency and transportation service levels. In the time window adjustment strategy, in order to coordinate the relationship between multiple transportation modes, a multi-mode coordination factor is introduced. The multimodal coordination factor is used to balance the cost and time factors of different transportation modes, and its calculation formula is: ; in, Indicates the path The unit transportation cost, Indicates the path The transportation time, and The maximum transportation cost and maximum transportation time among all the routes are respectively. Through standardization, the cost and time differences between different transportation modes are quantified into a comprehensive indicator The larger the index is, the higher the comprehensive transportation cost and time of the route is; otherwise, it is lower. The introduction of multimodal coordination factors can better balance the resource allocation between different modes of transportation in scheduling decisions and avoid excessive use or neglect of a certain mode of transportation. Integrate it into the optimization objective function to obtain a new optimization objective function: ; in, Indicates time period Slave Node Send to Node The decision variable of cargo volume is The penalty coefficient for goods that are not transported on time. and Respectively represent nodes The amount of cargo in the last time period and the initial time period. The first part of the optimization objective function It represents the minimization of the total transportation cost after introducing the multi-modal coordination factor, taking into account the cost and time factors of different paths. It is a penalty for failure to complete the transportation task in time, forcing the system to complete the transportation task as soon as possible while ensuring the lowest cost. In order to evaluate the effectiveness of the time window adjustment strategy, the effect evaluation index is designed. , and its calculation formula is ; in, and Respectively represent the weight coefficients of actual cost and actual time in the evaluation index. The actual cost and expected cost are the total cost of the current actual execution of the scheduling plan and the total cost predicted by the optimization model, respectively. The actual time and expected time are the total transportation time of the current actual Suxing and the total transportation time predicted by the model. When the value is greater than 1, it means that the actual execution cost and time are beyond expectations, and the scheduling strategy needs to be adjusted; When it is less than 1, it means that the actual implementation is better than expected, and the adjustment range should be appropriately reduced. The calculation result of is used to adaptively adjust the time window adjustment strategy for the next cycle. , it means that the current strategy cannot effectively meet the actual needs, and it is necessary to increase the time window adjustment strength of the next cycle, that is, increase the time window adjustment factor The value of When , it means that the current scheduling strategy is better than expected, and the adjustment intensity should be appropriately reduced to maintain the stability of the system and the flexibility of scheduling. The adaptive adjustment strategy can dynamically adjust the time window according to the actual operation situation, thereby improving the responsiveness and operation efficiency of the scheduling system. The constrained time window, the new optimization objective function and the adaptive adjustment strategy are organically combined to form a complete dynamic adjustment time window scheduling framework. The framework can flexibly adjust the time windows of each node and path in different time periods, and realize efficient allocation of resources between different transportation modes. The adjustment amplitude and direction of the time window are continuously optimized through the adaptive adjustment strategy.

[0030] In a specific embodiment, the process of executing step S5 may specifically include the following steps: Define the state space S and action space A to get the complete input space, where the state space contains the current cargo volume of all nodes, the current transportation volume of all paths, and the time window information, and the action space represents the transportation volume allocation decision for each path; Construct the Actor network fθ(s) and two Critic networks Qφ1(s,a) and Qφ2(s,a) to obtain the neural network structure, where θ and φ are network parameters respectively; Introducing the target network fθ' and Qφ1', Qφ2', we get the soft update mechanism, which uses the parameter update formula θ'←τθ+(1-τ)θ', φi'←τφi+(1-τ)φi', where τ is the soft update coefficient; Gaussian noise ε is added to the action generation process to obtain the exploration strategy μ(s). The calculation formula of the exploration strategy is μ(s)=π(s)+ε, where ε follows a normal distribution with a mean of 0 and a standard deviation of σ. Design an adaptive noise attenuation mechanism and obtain the noise standard deviation update formula σt=σ0*exp(-λt), where σ0 is the initial noise standard deviation, λ is the attenuation rate, and t is the number of training steps; Define the experience replay buffer D and get the sampling mechanism, which randomly extracts mini-batch data from D for training; Calculate the target Q value y=r+ω*min(Qφ1'(s',π(s')),Qφ2'(s',π(s'))) and get the network update formula. The update formula includes the Critic network update φi←φi-η∇φiL(φi) and the Actor network update θ←θ+η∇aQφ1(s,a)|a=π(s)∇θπ(s), ω represents the discount factor, and s' is the next state after executing action a; Design an adaptive learning rate adjustment mechanism and obtain the learning rate update formula η=η0 / (1+β*t), where η0 is the initial learning rate, β is the decay coefficient, and t is the number of training steps; The neural network structure, soft update mechanism, exploration strategy, noise standard deviation update formula, sampling mechanism, network update formula and learning rate update formula are combined to obtain a strategy generator.

[0031] Specifically, define the state space and action space , describing the current state of the multimodal transport system and possible decision operations. State space It contains the current cargo volume of all transportation nodes, the current transportation volume of all transportation paths, and time window information. The current cargo volume of each node indicates the amount of cargo waiting to be processed or transported at the node, reflecting the load of the node; the current transportation volume of each path indicates the transportation tasks currently scheduled on the path, reflecting the utilization of the path; the time window information contains the earliest shipment time and the latest arrival time of each node, which is used to determine the timeliness constraints of each batch of cargo during transportation. These state variables together constitute the state space of the system at a certain moment. , which can fully describe the current operation status of the entire transportation network. Action space It represents the transportation volume allocation decision for each path, that is, how much cargo should be transported from one node to another node through a certain path in the current state. By defining the state space and action space, a complete input space is constructed. Constructing the neural network structure in deep reinforcement learning, namely the Actor network And two critic networks and ,in and Represents the parameters of the network respectively. Actor Network The role is to receive the current state As input, and output an action , i.e. strategy , represents the optimal transport volume allocation decision critic network under the current state and is used to evaluate a given state-action pair The value, that is, the Q value, reflects the current state Next action The design of the dual critic network can effectively alleviate the overestimation bias problem caused by the single Q value network and improve the stability and convergence speed of the algorithm. In order to improve the stability and convergence effect of model training, the self-standard network is used for soft update mechanism. Introducing the target actor network and target critic network and , and their parameters are and The parameter update of the target network adopts a soft update mechanism, and its update formula is: in, is the soft update coefficient, which is usually small (such as 0.005) to ensure the slow change of the target network parameters. The soft update mechanism makes the parameters of the target network gradually approach the parameters of the actual network, thereby maintaining the stability of the target network during training and avoiding training instability caused by too fast parameter changes. In the process of strategy generation, in order to improve the exploration ability of the strategy, Gaussian noise is added when generating actions. , forming an exploration strategy The specific calculation formula is: ; in, The mean is 0 and the standard deviation is The normal distribution of By adding noise, the strategy generator is guided to explore more state space in the early stage of training to avoid falling into the local optimal solution too early. In order to gradually reduce exploration during the training process, an adaptive noise attenuation mechanism is designed, and the update formula of the noise standard deviation is: ; in, is the initial noise standard deviation, λ is the attenuation rate, is the number of training steps. As the training progresses, the noise standard deviation Gradually decreases, the strategy generator gradually shifts from the initial extensive exploration to the later precise strategy optimization. In order to effectively use the experience data, define the experience replay buffer , used to store the state-action-reward-next-state quadruple during training During each training, a mini-batch of data is randomly extracted from the experience pool for training to break the temporal correlation between samples and improve the efficiency and stability of training. The sampling mechanism of experience replay is random sampling to ensure the diversity of training samples, avoid the model's over-reliance on certain specific state-action pairs, and improve the generalization ability of the model. In each training, according to the current state and the actions generated by the strategy , and the reward obtained after performing actions in the environment and the next state , calculate the target Q value , and its calculation formula is: ; in, is the discount factor, usually taken as It is used to weigh the current reward and future reward. Indicates the current state Next action After that, the sum of the immediate reward and future long-term reward that can be obtained. By minimizing the target Q value With the current Q value and The mean square error between them is used to update the parameters of the Critic network. The update formula of the Critic network is: ; in, is the loss function, that is, the mean square error between the target Q value and the current Q value, is the learning rate. For the Actor network, the parameters of the policy network are updated by maximizing the Q value. The update formula is: ; By increasing the Q value corresponding to the output of the policy network in the Critic network, the quality of the policy is improved. In order to ensure the rationality of the learning rate at different training stages, an adaptive learning rate adjustment mechanism is designed. The learning rate update formula is: ; in, is the initial learning rate, is the learning rate decay coefficient, is the number of training steps. As the number of training steps increases, the learning rate Gradually decrease, so that the model can converge to the local optimal solution more stably in the later stage of training. The neural network structure, soft update mechanism, exploration strategy, noise standard deviation update formula, sampling mechanism, network update formula and learning rate update formula are combined to obtain the strategy generator.

[0032] In a specific embodiment, the process of executing step S6 may specifically include the following steps: Initialize the system parameters to obtain the initialization parameter set, which includes the network structure, time window and neural network weight. The network structure includes the number of nodes and path connection relationship, the time window includes the initial earliest shipment time and the latest arrival time, and the neural network weight includes the initial weight value of the Actor network and the Critic network. Input the current state into the demand forecaster to obtain the forecast value of freight demand and capacity supply for the next cycle. The current state includes the current cargo volume of each node, the current transportation volume of each path, and the current time window information. The dynamically adjusted time window is updated based on the predicted value to obtain an updated time window, wherein the updating process uses the time window adjustment formula in claim 5; Input the current state st into the strategy generator to obtain the scheduling decision at. The scheduling decision includes the transportation volume allocation of each path in each time period. The specific calculation process is at=μ(st)=π(st)+ε, where π(st) is the output of the Actor network and ε is the exploration noise. Based on the scheduling decision, the transportation process simulation is performed to obtain the actual transportation cost and time. The simulation process includes updating the cargo volume of each node and the transportation volume of each path, and calculating the total transportation cost and total transportation time; Calculate the reward value rt and obtain the reward function, which is rt=-F'(st,at), where F' is the new optimization objective function; Observe the new state st+1 after the decision is executed, and obtain the state transition sample (st, at, rt, st+1). The new state includes the updated cargo volume of each node, the transportation volume of each path, and the time window information; The state transition samples are stored in the experience playback buffer D to obtain an updated experience pool, the size of which is a preset value M; Randomly sample N mini-batch data from the updated experience pool to obtain a training sample set, where N is the preset batch size; The network parameters of the strategy generator are updated based on the training sample set to obtain the multimodal transport scheduling plan. The updating process includes: calculating the target Q value y=r+γ*min(Qφ1'(s',π(s')),Qφ2'(s',π(s'))), updating the Critic network parameters φi←φi-η∇φiL(φi), and updating the Actor network parameters θ←θ+η∇aQφ1(s,a)|a=π(s)∇θπ(s) every d steps, where η is the adaptive learning rate.

[0033] Specifically, the system parameters are initialized to obtain an initialization parameter set. The initialization parameter set includes the basic structure of the multimodal transport network, the time window setting, and the initial weight value of the neural network. The network structure describes the number of nodes and path connection relationships in the transportation system. For example, a multimodal transport system contains three main nodes: ports, railway hubs, and distribution centers. These three nodes are connected by roads, railways, and waterways to form a complete transportation network. The connection relationship of each path defines the accessibility and transportation mode between nodes. The initialization of the network structure can provide a basic framework for the entire scheduling optimization. Time window initialization includes the initial earliest shipping time of each node ( ) and the latest arrival time ( ). The neural network weight initialization includes the initial weight values ​​of the Actor network and the Critic network. The reasonable initialization of these weight values ​​can speed up the convergence of the network and improve the learning effect of the strategy generator. After completing the initialization of the system parameters, the current system state is input into the demand forecaster to predict the freight demand and capacity supply in the next cycle. The current state includes the current cargo volume of each transport node, the current transport volume of each transport path, and the current time window information. This information can reflect the real-time operating status of the system and provide effective input data for the demand forecaster. Based on historical data and current state information, the demand forecaster predicts the freight demand and capacity supply values ​​for the next cycle through the operation of a multi-layer feedforward neural network. Based on the prediction results of the demand forecaster, the dynamically adjusted time window is updated to obtain a more accurate time window setting. The update process uses the time window adjustment formula in claim 5, that is, by adjusting the factor Earliest shipping time and the latest arrival time Make adaptive adjustments. The amount of adjustment of the time window Based on forecast demand and current capacity The adjustment formula is: ; in, is the adjustment coefficient, which indicates the sensitivity to changes in demand. is the maximum adjustment amount, which is used to limit the amplitude of a single adjustment. Based on the calculated adjustment amount, the updated time window is: ; ; When the demand is predicted to increase, the shipment time can be appropriately advanced and the arrival time can be delayed, thereby alleviating the transportation pressure of the node and improving the flexibility of scheduling. Input strategy generator to generate scheduling decisions under the current state The policy generator learns and generates policies based on the deep deterministic policy gradient algorithm. The specific calculation process is: ; in, For the Actor network based on the current state The generated scheduling policy output, To explore noise, the mean is 0 and the standard deviation is Normal distribution This strategy generation method can guide the strategy generator to explore more state space in the early stage of training and avoid premature convergence to the local optimal solution. Including the transportation volume allocation of each path in each time period. Based on the generated scheduling decision , perform transportation process simulation in the transportation network and calculate the actual transportation cost and time. The transportation process simulation includes updating the cargo volume of each node and the transportation volume of each path, and calculating the total transportation cost and total transportation time. According to the scheduling decision, the transportation volume on each path is allocated to the corresponding node, and the transportation cost and time overhead caused by the scheduling plan are recorded. After the transportation simulation is completed, the reward value of the scheduling strategy is calculated , the reward function is in the form of: ; in, is the new optimization objective function, indicating that in the current state Execute scheduling decision The total cost and time caused by the scheduling decision. The negative sign indicates that the system hopes to minimize the total cost and time, that is, the larger the reward value, the better the system performance. The reward function is designed to guide the policy generator to generate a scheduling policy that can maximize the overall benefit of the system. After obtaining the current reward value, observe the new state after executing the current scheduling decision. , record the new status information, including the updated cargo volume of each node, the transportation volume of each path, and the time window information. , Scheduling Decision , Reward Value and the new state Composition state transition sample , and store it in the experience replay buffer , get the updated experience pool. The size of the experience pool is the preset value , used to store the experience data accumulated during the training process. Randomly sample from the updated experience pool The randomly sampled training sample set includes several state transition samples. These samples can provide rich state-action pairs and their corresponding reward information, thus supporting the training process of the policy generator. Based on the training sample set, the network parameters of the policy generator are updated, including the parameter updates of the Critic network and the Actor network. Calculate the target Q value , and its calculation formula is: ; in, is the discount factor, which indicates the degree of discount of future rewards, and They are the target Critic network, is the target Actor network. Target Q value Indicates the current state Next action After that, the sum of the immediate reward and future long-term reward that can be obtained. By minimizing the target Q value With the current Q value and The mean square error between them is used to update the parameters of the Critic network. The update formula is: ; in, is the loss function, that is, the mean square error between the target Q value and the current Q value, is the learning rate. For the Actor network, the parameters of the policy network are updated by maximizing the Q value. The update formula is: ; That is, by increasing the Q value of the output of the policy network in the Critic network, the quality of the policy is improved. The parameters of the Actor network are updated once every step, so that the Actor network can continuously optimize its strategy generation ability. During the training process, in order to maintain the rationality of the learning rate at different training stages, an adaptive learning rate adjustment mechanism is designed, and its learning rate update formula is: ; in, is the initial learning rate, is the learning rate decay coefficient, is the number of training steps. As the number of training steps increases, the learning rate gradually decreases, and the model can converge to the optimal solution more stably in the later stage of training. The updated results of the strategy generator are applied to the entire multimodal transport system to generate a specific scheduling plan. This scheduling plan can effectively respond to the dynamic changes in transportation demand based on multiple factors such as node cargo volume, path transportation volume, and time window, and provides an intelligent and dynamic scheduling optimization plan for the multimodal transport system.

[0034] The above describes the multimodal transport intelligent scheduling optimization method in the embodiment of the present invention. The following describes the multimodal transport intelligent scheduling optimization device in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an embodiment of the multimodal transport intelligent scheduling optimization device includes: A modeling module is used to model multiple transportation nodes and transportation paths to obtain a network topology structure; A partitioning module is used to divide the scheduling cycle into time periods based on the network topology to obtain a time window scheduling framework; A training module, used to input historical freight data into a multi-layer feedforward neural network for training to obtain a demand forecaster; An adjustment module, used for dynamically adjusting the time window of the time window scheduling framework to obtain a dynamically adjusted time window; A processing module, used for inputting the state space and the action space into the improved double-delayed deep deterministic policy gradient algorithm for training to obtain a policy generator; The decision module is used to generate multimodal transport scheduling decisions based on the demand forecaster, the dynamically adjusted time window and the strategy generator to obtain the multimodal transport scheduling plan.

[0035] Through the cooperation of the above components and the comprehensive modeling of multiple transport nodes and transport paths, a complete network topology is obtained. The scheduling cycle is divided into time periods based on the network topology, and a dynamic adjustment mechanism is introduced to enable the scheduling scheme to better adapt to changes in actual transportation demand. A multi-layer feedforward neural network is used to train historical freight data, and an adaptive adjustment mechanism is introduced to improve the accuracy of predicting future freight demand. The policy generator obtained by training with an improved double-delayed deep deterministic policy gradient algorithm can generate high-quality scheduling decisions in complex state spaces. By introducing multi-modal coordination factors and effect evaluation indicators, the balance between different modes of transportation is considered while optimizing the total transportation cost and time. Mechanisms such as dynamic time window adjustment and adaptive learning rate are adopted to enable the scheduling system to better adapt to environmental changes and uncertainties. Based on the demand predictor, the dynamically adjusted time window and the policy generator, the multimodal transport scheduling scheme can be generated and updated in real time, which improves the response speed of the system.

[0036] The present invention also provides a computer device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the multimodal transport intelligent scheduling optimization method in the above-mentioned embodiments.

[0037] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the multimodal transport intelligent scheduling optimization method.

[0038] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0039] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0040] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal transport intelligent scheduling optimization method, characterized in that: The method comprises: Model multiple transportation nodes and transportation paths to obtain the network topology; Divide the scheduling cycle into time periods based on the network topology to obtain a time window scheduling framework; The historical freight data is input into a multi-layer feed-forward neural network for training to obtain a demand forecaster; Dynamically adjusting the time window of the time window scheduling framework to obtain a dynamically adjusted time window; The state space and action space are input into the improved double-delayed deep deterministic policy gradient algorithm for training to obtain a policy generator; Based on the demand forecaster, the dynamically adjusted time window and the strategy generator, a multimodal transport scheduling decision is generated to obtain a multimodal transport scheduling solution.

2. The multimodal transport intelligent scheduling optimization method according to claim 1 is characterized in that: The multiple transport nodes and transport paths are modeled to obtain a network topology structure, including: Defining attributes of each transport node to obtain a node attribute set, wherein the node attribute set includes location coordinates, node type, and processing capacity index; Defining attributes for each transport route to obtain a set of route attributes, wherein the set of route attributes includes a starting point, an end point, a transport mode, a transport time, a transport cost, and a transport capacity upper limit; Based on the actual geographic information and transportation network data, the connection relationship between each node is determined to obtain a complete multimodal transport network topology. The historical freight data is collected and sorted to obtain the cargo inflow and outflow of each node and the actual transportation volume of each transportation route. An objective function is constructed based on the total transportation cost and the total transportation time to obtain a network optimization target, wherein the objective function is F=min(α*total transportation cost+β*total transportation time); Initialize the weight coefficients α and β in the objective function to obtain the initial weight value, and build network constraints based on the node attribute set and the path attribute set to obtain the initial network constraints; The network optimization target, the initial weight value and the initial network constraint are combined to obtain a network topology structure.

3. The multimodal transport intelligent scheduling optimization method according to claim 2 is characterized in that: The time window scheduling framework is obtained by dividing the scheduling period into time periods based on the network topology structure, including: The entire scheduling period T is evenly divided into K discrete time periods {t1, t2, ..., tK}, where K is the number of preset time periods; Define a state vector for each transport node Ni in each time period tk to obtain a node state vector Si,k, wherein the node state vector includes the current cargo volume Qi,k and the remaining processing capacity Ri,k; A state vector is defined for each transport route Rj in each time period tk to obtain a route state vector Pj,k, wherein the route state vector includes the current transport volume Vj,k and the remaining transport capacity Uj,k; Introduce decision variables xi,j,k to represent the amount of cargo shipped from node Ni to node Nj in time period tk, and obtain the set of scheduling decision variables; Based on the node capacity constraints, path capacity constraints and cargo balance constraints, the constraints are constructed to obtain the time window scheduling constraint set; For each batch of goods, the earliest shipping time TEi and the latest arrival time TLi are set, and the time window constraint TEi≤tk≤TLi-tj is obtained, where tj is the transportation time from node i to node j; The objective function is refined into F=min(∑i,j,k(cj*xi,j,k)+γ*∑imax(0,Qi,K-Qi,0)) to obtain the optimized objective function, where γ is the penalty coefficient for goods that are not transported on time; The discrete time period, the node state vector, the path state vector, the scheduling decision variable set, the time window scheduling constraint set, the time window constraint and the optimization objective function are combined to obtain a time window scheduling framework.

4. The multimodal transport intelligent scheduling optimization method according to claim 3 is characterized in that: The historical freight data is input into a multi-layer feedforward neural network for training to obtain a demand forecaster, including: Preprocessing the historical freight data to obtain standardized input features, wherein the standardized input features include historical freight volumes, seasonal factors, economic indicators, and weather conditions; Design a multi-layer feedforward neural network structure to obtain a neural network model, wherein the neural network model includes an input layer, multiple hidden layers and an output layer, wherein the hidden layer uses a ReLU activation function and the output layer uses a linear activation function; The standardized input features are divided into a training set and a validation set according to a ratio of 8:2, and the training set data is normalized to obtain normalized training data, wherein the normalization process scales all feature values ​​to the interval [0, 1]; Inputting the normalized training data into a neural network model, training using a back propagation algorithm and an Adam optimizer to obtain initial model parameters, wherein the training process uses a mean square error as a loss function; Applying the dropout technique to the initial model parameters to obtain model parameters that prevent overfitting, the dropout rate is set to 0.5; The model was evaluated using the k-fold cross-validation method to obtain model performance indicators, where the k value was set to 5, and the performance indicators included prediction accuracy, mean absolute error, and root mean square error; An adaptive adjustment mechanism is designed to obtain a demand forecaster. The adaptive adjustment mechanism includes: when the prediction error exceeds a preset threshold, model retraining is triggered, and the neural network weights are updated using an adaptive learning rate, wherein the adaptive learning rate η(t)=η0 / (1+λt), η0 is the initial learning rate, λ is the attenuation coefficient, t is the number of training steps, and η(t) is the adaptive learning rate.

5. The multimodal transport intelligent scheduling optimization method according to claim 4 is characterized in that: The dynamically adjusting the time window of the time window scheduling framework to obtain the dynamically adjusted time window includes: Define the time window adjustment factor δi,k, and obtain the adjustment amount calculation formula, which is δi,k=min(ρ*(Di,k-Ci,k),δmax), where Di,k is the predicted demand, Ci,k is the current capacity, ρ is the adjustment coefficient, and δmax is the maximum adjustment amount; The time window is updated based on the adjustment amount calculation formula to obtain an updated time window, and the update formula is TEi,k+1=TEi,k-max(0,δi,k), TLi,k+1=TLi,k+max(0,δi,k), where TEi,k is the earliest shipment time and TLi,k is the latest arrival time; The upper and lower limits of the updated time window are adjusted to obtain a constrained time window, wherein the constraint conditions are TEmin≤TEi,k≤TEmax, TLmin≤TLi,k≤TLmax; The multi-modal coordination factor μj is introduced to obtain the multi-modal balance formula, which is μj=(cj / cmax)*(tj / tmax), where cj is the transportation cost of path j, tj is the transportation time of path j, cmax and tmax are the maximum cost and maximum time of all paths respectively; Integrate the multi-mode coordination factor into the objective function to obtain a new optimization objective function F', wherein the new optimization objective function is F'=min(∑i,j,k(μj*cj*xi,j,k)+γ*∑imax(0,Qi,K-Qi,0)); Design effect evaluation index E, and obtain an evaluation formula, which is E=w1*(actual cost / expected cost)+w2*(actual time / expected time), where w1 and w2 are weight coefficients; Based on the calculation result of the evaluation formula, the time window adjustment strategy of the next cycle is adjusted to obtain an adaptive adjustment strategy. When E>1, the time window adjustment strength of the next cycle is increased, and when E<1, the adjustment amplitude of the next cycle is reduced; The constrained time window, the new optimization objective function and the adaptive adjustment strategy are combined to obtain a dynamically adjusted time window.

6. The multimodal transport intelligent scheduling optimization method according to claim 5 is characterized in that: The state space and action space are input into the improved double-delay deep deterministic policy gradient algorithm for training to obtain a policy generator, including: Define the state space S and action space A to get the complete input space, where the state space contains the current cargo volume of all nodes, the current transportation volume of all paths, and the time window information, and the action space represents the transportation volume allocation decision for each path; Construct the Actor network fθ(s) and two Critic networks Qφ1(s,a) and Qφ2(s,a) to obtain the neural network structure, where θ and φ are network parameters respectively; Introducing the target network fθ' and Qφ1', Qφ2', a soft update mechanism is obtained, wherein the soft update mechanism adopts a parameter update formula θ'←τθ+(1-τ)θ', φi'←τφi+(1-τ)φi', where τ is a soft update coefficient; Gaussian noise ε is added during the action generation process to obtain an exploration strategy μ(s). The calculation formula of the exploration strategy is μ(s)=π(s)+ε, where ε follows a normal distribution with a mean of 0 and a standard deviation of σ. Design an adaptive noise attenuation mechanism and obtain the noise standard deviation update formula σt=σ0*exp(-λt), where σ0 is the initial noise standard deviation, λ is the attenuation rate, and t is the number of training steps; Define an experience replay buffer D and obtain a sampling mechanism, which randomly extracts mini-batch data from D for training; Calculate the target Q value y=r+ω*min(Qφ1'(s',π(s')),Qφ2'(s',π(s'))) to get the network update formula, which includes the Critic network update φi←φi-η∇φiL(φi) and the Actor network update θ←θ+η∇aQφ1(s,a)|a=π(s)∇θπ(s), ω represents the discount factor, and s' is the next state after executing action a; Design an adaptive learning rate adjustment mechanism and obtain the learning rate update formula η=η0 / (1+β*t), where η0 is the initial learning rate, β is the decay coefficient, and t is the number of training steps; The neural network structure, the soft update mechanism, the exploration strategy, the noise standard deviation update formula, the sampling mechanism, the network update formula and the learning rate update formula are combined to obtain a strategy generator.

7. The multimodal transport intelligent scheduling optimization method according to claim 6 is characterized in that: The multimodal transport scheduling decision generation based on the demand forecaster, the dynamically adjusted time window and the strategy generator to obtain a multimodal transport scheduling solution includes: Initializing system parameters to obtain an initialization parameter set, wherein the initialization parameter set includes a network structure, a time window, and a neural network weight, wherein the network structure includes the number of nodes and path connection relationships, the time window includes an initial earliest shipping time and a latest arrival time, and the neural network weight includes initial weight values ​​of an Actor network and a Critic network; Input the current state into the demand forecaster to obtain the forecast value of freight demand and capacity supply for the next cycle, wherein the current state includes the current cargo volume of each node, the current transportation volume of each path and the current time window information; updating the dynamically adjusted time window based on the predicted value to obtain an updated time window, wherein the updating process uses the time window adjustment formula in claim 5; Input the current state st into the strategy generator to obtain the scheduling decision at, which includes the transportation volume allocation of each path in each time period. The specific calculation process is at=μ(st)=π(st)+ε, where π(st) is the output of the Actor network and ε is the exploration noise; Based on the scheduling decision, a transportation process simulation is performed to obtain actual transportation cost and time, wherein the simulation process includes updating the cargo volume of each node and the transportation volume of each path, and calculating the total transportation cost and total transportation time; Calculate the reward value rt to obtain a reward function, wherein the reward function is rt=-F'(st,at), where F' is a new optimization objective function; Observe the new state st+1 after the decision is executed, and obtain the state transition sample (st, at, rt, st+1), where the new state includes the updated cargo volume of each node, the transportation volume of each path, and the time window information; The state transition samples are stored in the experience playback buffer D to obtain an updated experience pool, where the size of the experience pool is a preset value M; Randomly sampling N mini-batch data from the updated experience pool to obtain a training sample set, where N is a preset batch size; The network parameters of the strategy generator are updated based on the training sample set to obtain a multimodal transport scheduling plan. The updating process includes: calculating the target Q value y=r+γ*min(Qφ1'(s',π(s')),Qφ2'(s',π(s'))), updating the Critic network parameters φi←φi-η∇φiL(φi), and updating the Actor network parameters θ←θ+η∇aQφ1(s,a)|a=π(s)∇θπ(s) every d steps, where η is the adaptive learning rate.

8. A multimodal transport intelligent dispatching optimization device, characterized in that: Used to execute the multimodal transport intelligent scheduling optimization method according to any one of claims 1 to 7, the multimodal transport intelligent scheduling optimization device comprises: A modeling module is used to model multiple transportation nodes and transportation paths to obtain a network topology structure; A division module, used to divide the scheduling period into time periods based on the network topology structure to obtain a time window scheduling framework; A training module, used to input historical freight data into a multi-layer feedforward neural network for training to obtain a demand forecaster; An adjustment module, used for dynamically adjusting the time window of the time window scheduling framework to obtain a dynamically adjusted time window; A processing module, used for inputting the state space and the action space into the improved double-delayed deep deterministic policy gradient algorithm for training to obtain a policy generator; A decision module is used to generate a multimodal transport scheduling decision based on the demand forecaster, the dynamically adjusted time window and the strategy generator to obtain a multimodal transport scheduling solution.

9. A computer device, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, it implements the multimodal transport intelligent scheduling optimization method described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the multimodal transport intelligent scheduling optimization method according to any one of claims 1 to 7.

Images