Artificial intelligence-based intelligent scheduling method for engineering vehicles

Abstract

The application discloses an engineering vehicle intelligent scheduling method based on artificial intelligence, comprising the following steps: constructing a scheduling target function, wherein the scheduling target function at least contains engineering vehicle transportation cost, task completion time cost, path congestion cost and energy consumption cost; based on historical task data and real-time traffic data, a multi-dimensional task demand prediction model is established to output task distribution prediction values and traffic state prediction values in a future period; the values are input into an engineering vehicle dynamic scheduling model based on a deep reinforcement learning algorithm for iterative optimization; based on the optimized dynamic scheduling model, real-time scheduling instructions are generated, engineering vehicle task paths and resources are dynamically allocated, and the execution state is monitored in real time to adjust the scheduling strategy. Through the construction of the scheduling target function, multi-dimensional demand prediction, deep reinforcement learning dynamic optimization and real-time scheduling and monitoring, the engineering vehicle intelligent scheduling is realized, which is efficient, economic and environmentally friendly.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of engineering vehicle management, and particularly relates to an engineering vehicle intelligent scheduling method based on artificial intelligence. BACKGROUND

[0002] With the acceleration of urbanization and the continuous expansion of infrastructure construction scale, the demand for engineering vehicles in engineering construction increases rapidly, which causes the problem of engineering vehicle fleet operation management efficiency, and especially the scheduling problem of engineering vehicles gradually becomes an important factor affecting the economy and execution efficiency of a project. Traditional engineering vehicle scheduling methods usually rely on manual experience scheduling methods, which are difficult to achieve global optimization and show obvious hysteresis and inefficiency in the face of complex and variable engineering task demands and real-time traffic conditions.

[0003] In recent years, with the development of artificial intelligence and machine learning technology, some emerging scheduling methods gradually introduce optimization algorithms to try to solve the above problems. For example, engineering vehicle path planning and task allocation based on simple historical data analysis or heuristic algorithms can improve local efficiency to a certain extent. However, this method still lacks deep and comprehensive consideration of multi-dimensional factors in complex scenarios, and cannot dynamically adapt to real-time changes in traffic conditions and task demands, so the overall optimization degree is still limited, and it is difficult to effectively balance the relationship between transportation cost, execution efficiency and energy consumption.

[0004] At present, the existing engineering vehicle scheduling technology generally has the following deficiencies: first, it lacks accurate prediction of future task demand and traffic state, and scheduling decisions are made only based on current information, resulting in strong task allocation blindness and easy occurrence of resource waste or task delay; second, the scheduling scheme ignores the comprehensive influence of transportation cost, path congestion cost and energy consumption, making it difficult to balance economy and environmental benefits; third, the existing technology fails to fully utilize real-time feedback data for dynamic adaptive adjustment, and the stability and robustness of the model are poor, which cannot effectively cope with real-time environmental changes, resulting in that the engineering vehicle scheduling decision cannot reach the global optimal state.

[0005] In summary, the existing engineering vehicle scheduling technology needs to be improved in terms of prediction accuracy, comprehensive cost control and dynamic decision optimization. Therefore, the existing engineering vehicle scheduling technology has the problems of low prediction accuracy of task demand and traffic state, insufficient comprehensive optimization degree and inability to dynamically adjust the scheduling strategy in real time. The present application discloses an engineering vehicle intelligent scheduling method based on artificial intelligence, which realizes dynamic allocation of engineering vehicle task path and resources by constructing a scheduling objective function considering transportation cost, task completion time cost, path congestion cost and energy consumption cost, and combining a multi-dimensional task demand prediction model and a dynamic scheduling model of deep reinforcement learning algorithm. SUMMARY

[0006] This section is intended to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments, and some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, the abstract and the title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0007] In view of the above existing problems, the present application is proposed.

[0008] To solve the above technical problems, the present application provides the following technical solutions: a scheduling target function is constructed, which at least includes engineering vehicle transportation cost, task completion time cost, path congestion cost and energy consumption cost;

[0009] Based on historical task data and real-time traffic data, a multi-dimensional task demand prediction model is established to output task distribution prediction values and traffic state prediction values in future time periods;

[0010] The task distribution prediction values and the traffic state prediction values are input into an engineering vehicle dynamic scheduling model established based on a deep reinforcement learning algorithm, the engineering vehicle dynamic scheduling model takes the scheduling target function as the optimization target, and iteratively optimizes the historical scheduling data and real-time feedback data;

[0011] Based on the optimized dynamic scheduling model, real-time scheduling instructions are generated, engineering vehicle task paths and resources are dynamically allocated, and the execution state is monitored in real time to adjust the scheduling strategy.

[0012] As a preferred scheme of the engineering vehicle intelligent scheduling method based on artificial intelligence according to the present application, the scheduling target function comprises:

[0013] Let f(x) represent the scheduling target function, and x represent the engineering vehicle scheduling scheme, then:

[0014] f(x) = α1·C trans (x) + α2·C time (x) + α3·C congestion (x) + α4·C energy (x)

[0015] Wherein, α1, α2, α3, α4 are adjustable weight coefficients, C trans (x) is the transportation cost, C time (x) is the task completion time cost, C congestion (x) is the path congestion cost, C energy (x) is the energy consumption cost.

[0016] As a preferred scheme of the engineering vehicle intelligent scheduling method based on artificial intelligence according to the present application, it specifically comprises:

[0017] The transportation cost is measured according to the number of vehicles, transportation distance and unit transportation mileage cost, the single vehicle distance cost is defined as β1, the number of vehicles is N, and the total transportation cost under the scheduling scheme x is represented as:

[0018]

[0019] Wherein, d i (x) represents the total driving distance of the i-th engineering vehicle performing the task;

[0020] The task completion time cost is comprehensively calculated according to the execution time, waiting time and default time of each task, and the delay penalty coefficient is defined as β2, then the task completion time cost under the scheduling scheme x is represented as:

[0021]

[0022] Wherein, M is the total number of tasks to be performed, T j (x) represents the total time consumption of the j-th task;

[0023] The path congestion cost is weighted calculated according to the congestion degree of different roads in different time periods, and the congestion cost is defined as β3, then the path congestion cost under the scheduling scheme x is represented as:

[0024]

[0025] Wherein, K is the total number of all available roads, ρ k is the congestion coefficient of road k, d k (x) is the driving distance of the engineering vehicle on road k;

[0026] The energy consumption cost comprehensively considers the fuel / electricity consumption of the vehicle, and the single vehicle energy consumption cost is defined as β4, then the energy consumption cost under the scheduling scheme x is represented as:

[0027]

[0028] Wherein, E i (x) represents the total energy consumption of the i-th engineering vehicle under the scheduling scheme x.

[0029] As a preferred scheme of the engineering vehicle intelligent scheduling method based on artificial intelligence, the multi-dimensional task demand prediction model comprises:

[0030] The historical task data at least contains time mark, geographical position, task type and vehicle use condition of the task;

[0031] The real-time traffic data at least includes road traffic, congestion index, traffic accident and restricted driving information;

[0032] The historical task data and the real-time traffic data are subjected to missing value filling, outlier removal and data format unification processing;

[0033] The historical task data is rearranged according to time sequence, and time sequence features of different periods of days, weeks and months are extracted;

[0034] An LSTM-based time sequence model is adopted to jointly predict short-term and medium-and long-term task demand trends, wherein the root mean square error of the time sequence model on a validation set is calculated as an evaluation index, if the evaluation index is higher than a first threshold value, model parameter optimization is automatically performed, otherwise the next step is entered;

[0035] According to the geographical position, the historical task data is subjected to grid processing to obtain a plurality of spatial sub-regions, for each sub-region, a spatial distribution prediction model is constructed in combination with task occurrence frequency and vehicle driving distance features, wherein in each sub-region, the deviation of the predicted value from the actual historical data is compared, if the deviation exceeds a second threshold value, the spatial distribution prediction model is retrained, otherwise the existing model is maintained;

[0036] Based on the real-time collected road traffic and congestion index, a regression model is adopted to dynamically predict the future congestion of different road sections, and traffic accidents and restricted driving information are superimposed to correct the congestion probability of each road section under different time windows, wherein if some road sections are in a high congestion state in the medium and long term prediction, they are marked as risk road sections and the congestion cost weight is increased, otherwise they are processed according to the conventional congestion coefficient;

[0037] According to the engineering vehicle task category, a classification model is used to independently predict each type of task demand, and the judgment results obtained in the above steps are multi-dimensionally fused to output task distribution prediction values and traffic state prediction values in the future period;

[0038] If the overall prediction error after fusion is greater than a third threshold value, iterative training is re-performed, otherwise the prediction result is output.

[0039] As a preferred scheme of the engineering vehicle intelligent scheduling method based on artificial intelligence, the engineering vehicle dynamic scheduling model comprises:

[0040] A deep Q network is selected as a deep reinforcement learning algorithm;

[0041] The task distribution prediction value and the traffic state prediction value are used as an environment state vector of the deep reinforcement learning algorithm;

[0042] defining an action space executable in the deep reinforcement learning algorithm, the action space comprising a dispatch assignment of the vehicle, a path selection;

[0043] reward splitting based on the dispatch objective function, wherein a positive reward is given if the dispatch decision reduces transportation cost, time cost, congestion cost and vehicle energy consumption, and a negative reward is given if the dispatch decision causes delay, congestion and excessive energy consumption;

[0044] cumulative rewards are used to update model parameters after the end of each dispatch decision cycle.

[0045] As a preferred scheme of the artificial intelligence-based intelligent scheduling method of engineering vehicles, the engineering vehicle dynamic scheduling model takes the dispatch objective function as an optimization objective and iteratively optimizes historical scheduling data and real-time feedback data, comprising:

[0046] initial policy network is obtained through preliminary training of historical scheduling data, wherein when it is found in the training process that the average reward of the model in the verification environment does not meet the fourth threshold value after continuous iterations for a certain number of times, the learning rate is adjusted, otherwise the training is continued until convergence;

[0047] The engineering vehicle dynamic scheduling model after preliminary training is deployed in the scheduling system, and a dispatch decision is obtained by inputting the environment state vector;

[0048] The objective function value under the dispatch decision is compared with the historical average value, if the overall cost significantly decreases, the next step is entered, otherwise the model iteration is continued;

[0049] The objective function value comprises transportation cost, time cost, congestion cost and energy consumption.

[0050] The running state of the engineering vehicle and the task execution completion are obtained in real time, and real-time scheduling feedback data is written into a data buffer pool for data updating;

[0051] The engineering vehicle dynamic scheduling model after preliminary training is fine-tuned online through the real-time scheduling feedback data;

[0052] If the average cumulative reward obtained by a new round of training is less than the fifth threshold value than the previous round, and this continues for a certain number of iterations, the network structure parameters are automatically adjusted, and if the average cumulative reward keeps increasing, the iteration training is continued;

[0053] The process is repeated until the optimal level of the cost target is reached.

[0054] As a preferred scheme of the intelligent scheduling method of the engineering vehicle based on artificial intelligence, a real-time scheduling instruction is generated based on the optimized dynamic scheduling model, and a task path and resources of the engineering vehicle are dynamically allocated, including:

[0055] A current optimal scheduling decision is obtained from the optimized dynamic scheduling model;

[0056] For each to-be-executed task, a scheduling instruction is generated in combination with the current state of the vehicle, and the scheduling instruction includes task start and end positions, a route to pass through, an estimated execution time, and required resource configuration; wherein the start and end times of each task, required mileage, and vehicle remaining load are quickly checked, if resources are insufficient or time conflicts occur, a replacement vehicle is called or the task is split, otherwise the check is passed;

[0057] Based on the traffic state prediction value, a path with low congestion risk and short driving distance is initially selected; wherein if the path along the line is detected to have a traffic restriction, a serious traffic accident, or serious congestion, a re-planning strategy is automatically triggered to select a suboptimal path, otherwise the original path is executed;

[0058] Considering the functional characteristics of the vehicle type and the fuel type, optimal matching is performed according to the task attributes, when a single vehicle function cannot meet the demand, the task is split and multi-vehicle collaborative scheduling is adopted, wherein if the additional cost of collaborative scheduling is higher than the expected income, the single vehicle scheduling scheme is returned, otherwise the multi-vehicle collaborative scheme is executed.

[0059] As a preferred scheme of the intelligent scheduling method of the engineering vehicle based on artificial intelligence, the execution state is monitored in real time to adjust the scheduling strategy, including:

[0060] The geographic coordinates of the vehicle are collected in real time by satellite positioning, and the running track of the vehicle during task execution is recorded;

[0061] The time consumption and completion progress of the vehicle in different stages of loading, driving, and unloading are monitored, and the task completion percentage is counted;

[0062] The vehicle load information and fuel / battery remaining amount information are collected, if it is detected that the vehicle energy consumption is high / the battery remaining amount is insufficient, a resource replenishment operation is triggered;

[0063] The actual execution time, energy consumption level, and path condition are compared with the estimated execution time, energy consumption level, and path condition in the scheduling instruction;

[0064] If the actual execution time exceeds the preset delay threshold, an emergency scheduling strategy is triggered;

[0065] If the actual energy consumption of the vehicle exceeds the preset energy consumption threshold, the record is fed back to the dynamic scheduling model as a negative sample;

[0066] If the execution result is basically consistent with the model prediction result, the execution result is taken as a positive sample to consolidate the model strategy.

[0067] As a preferred scheme of the intelligent scheduling method of the engineering vehicle based on artificial intelligence, the adjustment of the scheduling strategy comprises the following steps for dynamically correcting the engineering vehicle scheduling according to the monitoring result of the execution state:

[0068] When the vehicle encounters sudden congestion, accident or vehicle anomaly during execution, the dynamic scheduling model re-calculates the optimal path according to the new environment state vector; wherein if the reduction of the total cost after re-planning is greater than the switching cost caused by re-planning, the new scheduling scheme is executed, otherwise the existing strategy is maintained;

[0069] If it is monitored that multiple batches of tasks are delayed or the resource allocation is seriously unbalanced, the matching relationship between all available vehicles and tasks, the driving path and the execution order are re-calculated, and new global scheduling instructions are deployed.

[0070] The beneficial effects of the present application are:

[0071] 1. By constructing a scheduling objective function that comprehensively considers the transportation cost of the engineering vehicle, the task completion time cost, the path congestion cost and the energy consumption cost, the quantification and trade-off of various important cost factors in the engineering vehicle scheduling process are realized, a unified and comprehensive optimization target is provided for intelligent scheduling, the comprehensive cost is more effectively controlled in the actual scheduling process, and the economy and environmental friendliness of the engineering vehicle operation are improved;

[0072] 2. By establishing a task demand prediction model based on historical task data and real-time traffic data, including multiple dimensions of time, space, traffic state and task type, the precise prediction of future engineering task demand and traffic conditions is realized, the scheduling scheme of the engineering vehicle is more predictable and more forward-looking, the blindness and uncertainty of resource allocation in the task execution process are effectively reduced, and the risk of path congestion, task delay and resource waste is avoided in advance;

[0073] 3. By dynamically responding to and optimizing the complex and changeable real-time traffic and engineering task state, the learning ability of the model to the optimal decision under different scenes is continuously updated, in the actual scheduling process, the model can adaptively make better decisions and continuously optimize its scheduling strategy, and the robustness and stability of the scheduling model to external disturbances are improved;

[0074] 4. The intelligent scheduling strategy is implemented efficiently and accurately, and the actual effect is fed back to the dynamic scheduling model in real time. The scheduling strategy is continuously corrected and improved, further improving the efficiency of engineering vehicle resource allocation, the timeliness of task completion, the rationality of route planning and energy utilization efficiency. It effectively avoids real-time emergencies and task execution deviations that traditional static scheduling methods cannot handle. Attached Figure Description

[0075] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0076] Figure 1 This is a flowchart illustrating the intelligent scheduling method for engineering vehicles based on artificial intelligence as described in this invention. Detailed Implementation

[0077] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0078] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0079] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0080] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates an intelligent scheduling method for engineering vehicles based on artificial intelligence, which specifically includes the following steps:

[0081] S1. Construct a scheduling objective function, which should include at least the engineering vehicle transportation cost, task completion time cost, route congestion cost, and energy consumption cost. Note that the following points should be noted in this step:

[0082] Transportation costs are measured based on the number of vehicles, transportation distance, and cost per unit distance. Let β1 be the cost per vehicle distance and N be the number of vehicles. Then, the total transportation cost under scheduling scheme x is expressed as:

[0083]

[0084] wherein d i (x) represents the total mileage of the i-th engineering vehicle performing the task;

[0085] The task completion time cost is comprehensively calculated according to the execution time, waiting time and default time of each task, and the delay penalty coefficient is defined as β2, and the task completion time cost under the scheduling scheme x is represented as:

[0086]

[0087] wherein M is the total number of tasks to be performed, T j (x) represents the total time consumption of the j-th task;

[0088] The path congestion cost is weighted calculated according to the congestion degree of different roads in different time periods, and the congestion cost is defined as β3, and the path congestion cost under the scheduling scheme x is represented as:

[0089]

[0090] wherein K is the total number of all available roads, ρ k is the congestion coefficient of road k, d k (x) is the driving distance of the engineering vehicle on road k;

[0091] The energy consumption cost comprehensively considers the fuel / electricity consumption of the vehicle, and the energy consumption cost of a single vehicle is defined as β4, and the energy consumption cost under the scheduling scheme x is represented as:

[0092]

[0093] wherein E i (x) represents the total energy consumption of the i-th engineering vehicle under the scheduling scheme x.

[0094] As an example, the scheduling objective function is represented by the following formula, wherein f(x) represents the scheduling objective function, x represents the engineering vehicle scheduling scheme, and then:

[0095] f(x) = α1·C trans (x) + α2·C time (x) + α3·C congestion (x) + α4·C energy (x)

[0096] wherein α1, α2, α3, α4 are adjustable weight coefficients, C trans (x) is the transportation cost, C time (x) is the task completion time cost, C congestion (x) is the path congestion cost, Cenergy (x) is the energy consumption cost.

[0097] S2, based on historical task data and real-time traffic data, a multi-dimensional task demand prediction model is established, and the task distribution prediction value and traffic state prediction value in the future period are output. It needs to be explained that this step is:

[0098] The historical task data at least contains the time mark (such as date, hour), geographical position (such as latitude and longitude, area number), task type and vehicle usage (such as vehicle number, vehicle load, scheduling frequency) of the task;

[0099] The real-time traffic data at least contains road traffic, congestion index, traffic accident and traffic restriction information;

[0100] The historical task data and real-time traffic data are processed for missing value filling, outlier removal and data format unification;

[0101] The historical task data is rearranged according to time sequence, and the time sequence characteristics of different periods such as day, week and month are extracted;

[0102] The time sequence model based on LSTM is adopted to jointly predict the short-term and long-term task demand trend, wherein the root mean square error of the time sequence model on the validation set is calculated as an evaluation index, if the evaluation index is higher than the first threshold value, the model parameter optimization is automatically executed, otherwise the next step is entered;

[0103] Exemplarily, let X t represent the historical time sequence data at time t, represent the demand prediction value of the model at time t and the subsequent period;

[0104] The long short-term memory (LSTM) network is adopted to model X t , and output

[0105] The root mean square error RMSE of the model on the validation set is evaluated, if:

[0106]

[0107] The model parameter optimization is automatically executed;

[0108] If RMSE≤TH1, the next step is entered;

[0109] The first threshold value (TH1) is set to 1.2-1.5 times of the average RMSE on the past training set, if it exceeds the proportion, it indicates that the model prediction deviation is too large, and the learning rate needs to be adjusted.

[0110] In an optional embodiment, the historical task data is divided into a plurality of spatial sub-regions according to geographical location, each sub-region corresponding to a road region in the city;

[0111] For each sub-region, the historical task frequency, vehicle travel distance and vehicle demand characteristics are counted;

[0112] Let F z represent the historical feature vector of the sub-region z (such as the task frequency and vehicle travel distance average in the past period), represent the task demand prediction value of the sub-region in the future period;

[0113] A multiple linear regression model is used to estimate the demand of each sub-region;

[0114] The deviation of the prediction value from the actual historical data D z is calculated. If Δ z > the second threshold value (TH2), it indicates that the model has poor prediction accuracy in this sub-region and needs to be retrained;

[0115] If Δ z ≤ TH2, the existing model is maintained;

[0116] The second threshold value (TH2) is determined based on the importance of each sub-region to the overall scheduling decision.

[0117] In an optional embodiment, let G k,t represent the traffic feature vector of road k collected at time t (including road traffic, congestion index, whether an accident or traffic restriction occurs), represent the congestion degree prediction value of road k in the future period t;

[0118] A random forest regression model is used to estimate ;

[0119] If some roads k appear for a long-term period in the prediction, which exceeds the preset congestion critical value, the road k is marked as a high-risk congestion road segment, and its congestion cost weight is increased in the subsequent scheduling decision;

[0120] If remains at a normal level, it is treated as a normal congestion coefficient;

[0121] When the system detects that road k has an accident or has a temporary traffic restriction announcement, the is corrected, and a more actual congestion prediction value is output.

[0122] Further, according to the engineering vehicle task category, the classification model is used to independently predict the task demand of each category, multi-dimensional fusion is performed on the judgment results obtained in the above steps, and task distribution prediction values and traffic state prediction values in a future period are output.

[0123] If the overall prediction error after fusion is greater than the third threshold value, iterative training is performed again, otherwise the prediction result is output, that is, the task distribution prediction value and the traffic state prediction value in the future period are output.

[0124] The engineering vehicle task category at least includes earthwork transportation, equipment transportation, and concrete distribution.

[0125] As an example, the multi-dimensional task demand prediction model is represented as:

[0126]

[0127] Wherein, X t is a high-dimensional input vector integrating time, space, traffic, and task type information, represents a task distribution prediction value in a future period t, represents a traffic state prediction value in a future period t.

[0128] It should be noted that the time dimension prediction and the space dimension prediction are merged by weighting to obtain traffic state prediction and task type prediction are spliced in the feature space to obtain comprehensive prediction of different types of task demand under road congestion.

[0129] Finally, the above multi-dimensional results are integrated into a unified prediction output:

[0130]

[0131] The comprehensive error index (such as multi-dimensional RMSE weighted) is selected to evaluate the fusion and ;

[0132] If the comprehensive error is greater than the third threshold value TH3, iterative training is performed again.

[0133] If the comprehensive error is less than or equal to the third threshold value TH3, the prediction result is output as the input of the subsequent dynamic scheduling model.

[0134] S3, input the task distribution prediction value and the traffic state prediction value into an engineering vehicle dynamic scheduling model established based on a deep reinforcement learning algorithm, the engineering vehicle dynamic scheduling model taking a scheduling target function as an optimization target, and iteratively optimizing based on historical scheduling data and real-time feedback data. It should be noted that the present step is as follows:

[0135] Let π θ represent a decision strategy function of the engineering vehicle dynamic scheduling model, wherein θ is a learnable parameter of the model, and in each decision cycle, the model is based on the current environment state s t to output a scheduling allocation and path selection action a t for the vehicle.

[0136] The environment state s t is composed of the task distribution prediction value and the traffic state prediction value output by step S2.

[0137] In the DQN (Deep Q Network), a function Q(s, a; θ) is defined to estimate the expected return value that can be obtained by taking action a in state s.

[0138] The value of θ is constantly updated through backpropagation, so that the estimation of Q(s, a; θ) approximates the true return.

[0139] In an optional embodiment, the scheduling target function f(x) constructed in the present embodiment step S1, wherein x represents an engineering vehicle scheduling scheme, the function at least contains transportation cost, task completion time cost, path congestion cost and energy consumption cost.

[0140] When the model successfully reduces one or more of the above cost factors in one scheduling decision, a positive reward is given; if it causes delay, congestion or excessive energy consumption, a negative reward is given.

[0141] It should be noted that the reward value r t is calculated by comparing the improvement of the present scheduling decision in several cost indicators.

[0142] Let A be the action space in the present invention, each scheduling decision contains task allocation selection and path selection for several engineering vehicles.

[0143] In the DQN, the current best action is determined by max a∈A Q(s t , a; θ) in order to maximize the cumulative reward in the long run.

[0144] The historical scheduling data (such as the real recorded vehicle scheduling decisions, cost spending, traffic congestion degree in the past projects) is constructed into a training sample set in the format of 〈s, a, r, s'〉.

[0145] The data is used as the experience replay pool for offline training of DQN, helping the model learn the initial Q function;

[0146] Randomly sample small batches of samples from the training sample set, and iteratively update Q(s, a; θ);

[0147] After several rounds of training, a preliminary policy network is obtained The average reward value on the validation set is denoted as

[0148] Let TH4 (the fourth threshold) represent the minimum requirement for the average reward in the validation environment. If the average reward is lower than TH4 for consecutive iterations, The learning rate is adjusted; if The model has reached a usable level, and the next stage of deployment and online training is entered;

[0149] The setting of TH4 is based on the average cost level of historical decisions.

[0150] In an optional implementation, the is deployed in the engineering vehicle dispatching system;

[0151] In each decision-making cycle, the latest environment state s t is input, and the model outputs the action a t ;

[0152] If the objective function value (such as the comprehensive cost) obtained after the current dispatching decision is significantly lower than the historical average, it indicates that the decision has improved significantly; otherwise, the model needs to continue iteration;

[0153] The system continuously obtains vehicle operating status and task execution completion information, and adds the newly generated t , t , t , t+1 to the data buffer pool for subsequent online training;

[0154] Let represent the average reward in the nth round of online training phase;

[0155] At the end of each cycle, according to the real-time dispatching feedback data, the θ is updated in small batches, so that Q(s, a; θ) gradually adapts to the new environment state and task demand;

[0156] Let TH5 (the fifth threshold) represent the average reward improvement threshold of the model in consecutive iterations, which is used to determine whether the model continues to maintain effective improvement;

[0157] If the obtained in the new round of training is relatively lower than the last round the promotion If the improvement of the model is slow and there is no significant improvement after multiple iterations, automatically perform network structure parameter adjustment (such as adding or deleting network layers);

[0158] If If it is still greater than TH5, it means that the model is still in the rapid convergence stage, and no large-scale structure adjustment is needed;

[0159] After multiple iterations of online training, if If it is less than TH5, it means that the model tends to converge and a relatively mature engineering vehicle dynamic scheduling strategy is obtained;

[0160] Let θ * The final dynamic scheduling model is denoted as

[0161] Exemplarily, Deployed on the weights θ * of the deep neural network, when receiving the environment state s t , the model outputs the optimal action a

[0162]

[0163] Thus, the dynamic allocation of tasks and path planning for engineering vehicles is realized.

[0164] S4, based on the optimized dynamic scheduling model, generate real-time scheduling instructions, dynamically allocate engineering vehicle task paths and resources, and real-time monitor the execution state to adjust the scheduling strategy. It needs to be explained that in this step:

[0165] Obtain the current optimal scheduling decision from the optimized dynamic scheduling model;

[0166] For each to-be-executed task, generate a scheduling instruction combining the current state of the vehicle. The scheduling instruction includes the task start and end position, the route, the expected execution time, and the required resource configuration. For each task, the start and end time, the required mileage, and the remaining vehicle load are quickly checked. If the resources are insufficient or the time is in conflict, a substitute vehicle is called or the task is split, otherwise the check is passed.

[0167] Based on the traffic state prediction value, initially select a path with low congestion risk and short driving distance. If a limit line, serious traffic accident, or serious congestion is detected along the path, automatically trigger the re-planning strategy to select a suboptimal path, otherwise execute the original path.

[0168] For example, if Δf(x) > Δcost switch , a path switch is performed to switch to the sub-optimal path.

[0169] where Δf(x) represents the overall cost reduction due to the change in path, and Δcost switch represents the additional burden caused by the path switch.

[0170] Considering the functional characteristics of the vehicle type and the fuel type, the optimal matching is performed according to the task attributes, and when a single vehicle function cannot meet the demand, the task is split and multi-vehicle collaborative scheduling is adopted, wherein if the additional cost of collaborative scheduling is higher than the expected benefit, the single vehicle scheduling scheme is returned, otherwise the multi-vehicle collaborative scheme is executed.

[0171] For example, if ΔR co-op > ΔC coordination , the multi-vehicle collaborative scheme is executed.

[0172] where ΔR co-op represents the overall benefit increment after multi-vehicle collaboration, and ΔC coordination represents the additional cost of vehicle fleet collaborative scheduling.

[0173] In an optional embodiment, the execution state is monitored in real time to adjust the scheduling strategy, including:

[0174] The geographic coordinates of the vehicle are collected in real time by satellite positioning, and the running track of the vehicle during the execution of the task is recorded.

[0175] The time consumption and completion progress of the vehicle in different stages of loading, driving and unloading are monitored, and the task completion percentage is counted.

[0176] The vehicle load information and fuel / battery remaining information are collected, and if it is detected that the vehicle energy consumption is high / the battery remaining amount is insufficient, a resource replenishment operation is triggered.

[0177] The actual execution time, energy consumption level and path situation are compared with the estimated execution time, energy consumption level and path situation in the scheduling instruction.

[0178] If the actual execution time exceeds the preset time delay threshold, an emergency scheduling strategy is triggered.

[0179] If the actual energy consumption of the vehicle exceeds the preset energy consumption threshold, the record is fed back to the dynamic scheduling model as a negative sample.

[0180] If the execution result is basically consistent with the model prediction result, it is regarded as a positive sample to consolidate the model strategy.

[0181] Specifically, the adjustment of the scheduling strategy includes the following steps for dynamically correcting the engineering vehicle scheduling according to the monitoring result of the execution state:

[0182] When the vehicle encounters sudden congestion, accidents or vehicle abnormalities during the execution process, the optimal path is recalculated according to the new environmental state vector by the dynamic scheduling model; wherein, if the reduction of the total cost after re-planning is greater than the switching cost brought by re-planning, the new scheduling scheme is executed, otherwise the existing strategy is maintained;

[0183] If it is monitored that multiple batches of tasks are delayed or the resource allocation is seriously unbalanced, the matching relationship, driving path and execution order of all available vehicles and tasks are recalculated, and new global scheduling instructions are deployed.

[0184] The foregoing data missing value filling, outlier elimination and data format unification processing method can be realized by using the existing technical means, which will not be described in detail in this example.

[0185] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. An intelligent scheduling method for engineering vehicles based on artificial intelligence, characterized in that, include: Construct a scheduling objective function, which includes at least the engineering vehicle transportation cost, task completion time cost, route congestion cost, and energy consumption cost; Based on historical task data and real-time traffic data, a multi-dimensional task demand prediction model is established to output predicted task distribution and traffic conditions for future time periods. The multi-dimensional task demand prediction model includes: The historical mission data shall include at least the mission's timestamp, geographical location, mission type, and vehicle usage. The real-time traffic data includes at least road traffic flow, congestion index, traffic accidents, and traffic restriction information; The historical task data and the real-time traffic data are processed by filling in missing values, removing outliers, and unifying data formats. The historical task data is rearranged according to time series, and the temporal features of different periods of day, week, and month are extracted; An LSTM-based time series model is used to jointly predict the short-term and medium-to-long-term task demand trends. The root mean square error of the time series model on the validation set is used as an evaluation index. If the evaluation index is higher than a first threshold, the model parameters are automatically tuned; otherwise, the process proceeds to the next step. Based on the geographical location, the historical task data is processed into a grid to obtain several spatial sub-regions. For each sub-region, a spatial distribution prediction model is constructed by combining the frequency of task occurrence and vehicle travel distance characteristics. Within each sub-region, the deviation between the predicted value and the actual historical data is compared. If the deviation exceeds a second threshold, the spatial distribution prediction model is retrained; otherwise, the existing model is maintained. Based on real-time collected road traffic flow and congestion index, a regression model is used to dynamically predict the future congestion of different road sections. Traffic accident and traffic restriction information are superimposed to correct the congestion probability of each road section under different time windows. If some road sections are in a high-congestion state for a long time in the prediction, they are marked as risk road sections and the congestion cost weight is increased; otherwise, they are treated as regular congestion coefficients. Based on the engineering vehicle task category, the classification model is used to independently predict the demand for each type of task. The judgment results obtained from the above steps are fused in multiple dimensions to output the task distribution prediction value and traffic state prediction value in the future time period. If the overall prediction error after fusion is greater than the third threshold, then iterative training is performed again; otherwise, the prediction result is output. The predicted task distribution and the predicted traffic state are input into a dynamic scheduling model for engineering vehicles established based on a deep reinforcement learning algorithm. The dynamic scheduling model uses the scheduling objective function as its optimization objective and iteratively optimizes it using historical scheduling data and real-time feedback data. The dynamic scheduling model for engineering vehicles includes: Deep Q-networks were selected as the deep reinforcement learning algorithm. The predicted task distribution and the predicted traffic state are used as the environment state vector for the deep reinforcement learning algorithm. Define the executable action space in the deep reinforcement learning algorithm, the action space including vehicle scheduling and allocation, and path selection; The reward is split based on the scheduling objective function. If the scheduling decision reduces transportation costs, time costs, congestion costs, and vehicle energy consumption, a positive reward is given. If the scheduling decision causes delays, congestion, and excessive energy consumption, a negative reward is given. At the end of each scheduling decision cycle, the accumulated reward is used to update the model parameters; Based on the optimized dynamic scheduling model, real-time scheduling instructions are generated, engineering vehicle task paths and resources are dynamically allocated, and the execution status is monitored in real time to adjust the scheduling strategy.

2. The intelligent scheduling method for engineering vehicles based on artificial intelligence according to claim 1, characterized in that, The scheduling objective function includes: make Represents the scheduling objective function. To represent the engineering vehicle dispatching plan, then: in, , , , For adjustable weighting coefficients, For transportation costs, For the time cost of completing the task, For route congestion costs, Energy consumption costs.

3. The intelligent scheduling method for engineering vehicles based on artificial intelligence according to claim 2, characterized in that, Specifically, it includes: The transportation cost is measured based on the number of vehicles, transportation distance, and cost per unit distance, with the cost per vehicle distance defined as... If the number of vehicles is N, then in the scheduling scheme The total transportation cost is expressed as follows: in, This represents the total mileage traveled by the i-th engineering vehicle during its mission. The task completion time cost is calculated comprehensively based on the execution time, waiting time, and default time of each task, and the delay penalty coefficient is defined as follows: In the scheduling scheme The time cost for completing the task is expressed as follows: Where M is the total number of tasks to be executed. This represents the total time consumed by the j-th task; The path congestion cost is calculated by weighting the congestion levels of different roads at different times, and the congestion cost is defined as follows: In the scheduling scheme The path congestion cost is expressed as: Where K is the total number of all usable roads. Let k be the congestion coefficient of road k. Let k be the distance the engineering vehicle travels on road k. The energy consumption cost, taking into account both vehicle fuel and electricity consumption, is defined as the energy cost per vehicle. In the scheduling scheme The energy consumption cost is expressed as follows: in, Indicates the i-th engineering vehicle in the dispatching scheme Total energy consumption.

4. The intelligent scheduling method for engineering vehicles based on artificial intelligence according to claim 1, characterized in that, The dynamic scheduling model for engineering vehicles uses the scheduling objective function as the optimization target and iteratively optimizes it using historical scheduling data and real-time feedback data, including: The initial policy network is obtained through preliminary training using historical scheduling data. If, during the training process, it is found that the average reward of the model in the validation environment does not meet the fourth threshold after several consecutive iterations, the learning rate is adjusted; otherwise, training continues until convergence. The preliminarily trained dynamic scheduling model for engineering vehicles is deployed in the scheduling system, and scheduling decisions are obtained by inputting the environmental state vector. The objective function value under the scheduling decision is compared with the historical average. If the overall cost decreases significantly, proceed to the next step; otherwise, continue model iteration. The objective function values ​​include transportation costs, time costs, congestion costs, and energy consumption. The system acquires the real-time operating status of engineering vehicles and the completion status of task execution, and writes the real-time scheduling feedback data into the data buffer pool for data updates. The dynamic scheduling model for engineering vehicles after initial training is fine-tuned online using the real-time scheduling feedback data. If the average cumulative reward obtained in the new round of training increases by less than the fifth threshold compared to the previous round, and this continues for several iterations, the network structure parameters will be automatically adjusted. If the average cumulative reward continues to increase, the training will continue iteratively. Repeat this process until the optimal cost target is achieved.

5. The intelligent scheduling method for engineering vehicles based on artificial intelligence according to claim 4, characterized in that, Based on the optimized dynamic scheduling model, real-time scheduling instructions are generated to dynamically allocate task paths and resources for engineering vehicles, including: Obtain the current optimal scheduling decision from the optimized dynamic scheduling model; For each task to be executed, a scheduling instruction is generated based on the current status of the vehicle. The scheduling instruction includes the task start and end locations, route, estimated execution time, and required resource configuration. The start and end times, required mileage, and remaining vehicle load of each task are quickly verified. If resources are insufficient or there is a time conflict, a backup vehicle is called or the task is split. Otherwise, the verification passes. Based on traffic condition predictions, routes with low congestion risk and short travel distances are initially selected. If traffic restrictions, serious traffic accidents, or severe congestion are detected along the route, a replanning strategy is automatically triggered to select a suboptimal route; otherwise, the original route is executed. Considering the functional characteristics of vehicle types and fuel types, optimal matching is performed based on task attributes. When a single vehicle function cannot meet the demand, the task is split and multi-vehicle collaborative scheduling is adopted. If the additional cost of collaborative scheduling is higher than the expected benefit, it will fall back to the single-vehicle scheduling scheme; otherwise, the multi-vehicle collaborative scheme will be executed.

6. The intelligent scheduling method for engineering vehicles based on artificial intelligence according to claim 5, characterized in that, Real-time monitoring of execution status to adjust scheduling strategies, including: The vehicle's geographic coordinates are collected in real time via satellite positioning, and the vehicle's trajectory during mission execution is recorded. Monitor the time consumption and progress of the vehicle at different stages of loading, driving, and unloading, and calculate the percentage of task completion; Collect vehicle load information and fuel / battery balance information. If high vehicle energy consumption or insufficient battery balance is detected, trigger resource replenishment operation. Compare the execution time, energy consumption level, and path information with the estimated values ​​in the scheduling instructions; If the measured execution time exceeds the preset delay threshold, an emergency scheduling strategy will be triggered. If the actual energy consumption of a vehicle exceeds the preset energy consumption threshold, the record will be fed back to the dynamic scheduling model as a negative sample. If the execution result is basically consistent with the model prediction result, it is used as a positive sample to consolidate the model strategy.

7. The intelligent scheduling method for engineering vehicles based on artificial intelligence according to claim 6, characterized in that, The adjustment of the scheduling strategy includes the following steps, used to dynamically correct the scheduling of engineering vehicles based on the monitoring results of the execution status: When a vehicle encounters sudden congestion, an accident, or a vehicle malfunction during its execution, the dynamic scheduling model recalculates the optimal path based on the new environmental state vector. If the reduction in total cost after replanning is greater than the switching cost brought about by replanning, the new scheduling scheme is executed; otherwise, the existing strategy is maintained. If multiple batches of tasks are detected to be delayed or resource allocation is severely unbalanced, the matching relationship between all available vehicles and each task, the driving path and execution order will be recalculated, and new global scheduling instructions will be deployed.

Images