Multi-vehicle-type trajectory collaborative prediction method based on Rhodeus deus optimized hierarchical game

Abstract

The invention discloses a multi-vehicle-type trajectory collaborative prediction method based on Rhodeus deus optimized hierarchical game, and the method comprises the steps: S1, constructing a multi-vehicle-type initial three-dimensional data matrix, carrying out the interpolation compensation and outlier elimination, and carrying out the unified coding and indexing; differential normalization processing is carried out; s2, constructing a layered skeleton network comprising a basic dynamics layer and a disturbance behavior layer, and obtaining preliminary prediction; s3, establishing a multi-vehicle-type coupling objective function, scene penalty and a dynamic radius mechanism, and outputting a global solution by adopting a rhodeus deus deus algorithm; s4, performing hierarchical game residual error correction on the global optimal trajectory to obtain a hierarchical game residual error; s5, coupling the residual error with a rhodeus deus algorithm, synchronously and finely adjusting the trajectory, and outputting an optimal multi-vehicle type prediction trajectory; and S6, applying the optimal multi-vehicle-type prediction trajectory to short-term trajectory prediction deployment of the autonomous vehicle, completing comprehensive prediction and scheduling of different vehicle types, and realizing high-precision trajectory prediction in a multi-vehicle-type mixed driving scene.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle trajectory prediction, and in particular to a multi-vehicle trajectory collaborative prediction method based on minnow optimization layered game. Background Art

[0002] The current multi-model traffic prediction and analysis solutions have achieved time series modeling and statistical regression of vehicle driving data to a certain extent. However, due to the different power characteristics and driving conditions of various models, the existing technology can only perform relatively single-dimensional extraction and classification when dealing with the acceleration differences between pure electric vehicles and fuel vehicles, and the high-frequency start-stop characteristics of public transportation vehicles. This makes it difficult to fully take into account the differences between individuals and the real-time nature of traffic as a whole when facing large-scale, dynamically evolving vehicle groups. Although existing studies have used map information and distributed computing, they often lack a highly interactive architecture for multi-model simultaneous prediction and joint optimization under complex and busy road conditions, making it difficult to accurately capture the balance between local sudden conditions and global stable situations.

[0003] In many solutions, the frequent stops of public transport vehicles are not included in the special disturbance modeling, and their impact is only simply handled in the overall optimization or statistical analysis, resulting in insufficient deep integration of data streams of official vehicles, private cars and buses, especially in dynamic changes and local abnormal conditions. Even if some solutions have initially connected data from multiple models, they lack detailed regulation of abnormal phenomena or sudden changes in working conditions in the later stages, making it difficult to achieve refined prediction and management of multiple vehicle types under a unified framework.

[0004] Therefore, it is urgent to design a multi-vehicle trajectory collaborative prediction method based on minnow optimization hierarchical game to solve the problems existing in the above-mentioned existing technologies. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-vehicle trajectory collaborative prediction method based on minnow optimization hierarchical game, aiming to achieve collaborative prediction and subsequent optimization of multi-type vehicle trajectories in urban traffic systems.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: A first aspect of the present invention provides a method for collaboratively predicting multi-vehicle trajectories based on minnow optimization hierarchical game, the method comprising the following steps: S1: construct an initial three-dimensional data matrix of multiple vehicle models, perform interpolation compensation on the null value records in the initial three-dimensional data matrix of multiple vehicle models, remove outliers exceeding the threshold, and uniformly encode and index different vehicle models; perform differentiated normalization processing on the feature dimension based on the multi-vehicle collaborative optimization scaling function to obtain a multi-vehicle feature vector sequence; S2: Based on the multi-vehicle type feature vector sequence, perform multi-vehicle type feature embedding and construction of a high-order coupling network layer, then construct a hierarchical skeleton network including a basic dynamics layer and a perturbation behavior layer to capture both regular acceleration / deceleration and short-term strong changes, and finally fuse the outputs of the two levels to obtain a preliminary prediction result; S3: Combine the preliminary prediction result with the vehicle prior features, establish a multi-vehicle type coupling objective function to form an initial solution set; construct a vehicle type scenario-based penalty and dynamic radius mechanism, and use an improved bitterling algorithm for parallel optimization to obtain the global optimal multi-vehicle type trajectory; S4: Perform hierarchical decomposition on the global optimal trajectory and construct basic dynamics; construct a residual game mechanism and iteratively update the perturbation correction layer, and combine the two-layer residuals to obtain hierarchical game residuals; S5: Couple the hierarchical game residuals with the search process of the improved bitterling algorithm, synchronously fine-tune the multi-vehicle type trajectory parameters and residual information, and output the optimal multi-vehicle type prediction trajectory; S6: Apply the optimal multi-vehicle type prediction trajectory to the short-term trajectory prediction deployment of autonomous vehicles to complete the comprehensive prediction and scheduling of different vehicle types.

[0007] As an embodiment of the present application, step S1 specifically includes: S11: Obtain multi-dimensional information such as power output, acceleration, vehicle start / stop times, and historical trajectories of battery electric vehicles, fuel vehicles, and public transportation vehicles, organize the above multi-dimensional information in a time series, and construct a multi-vehicle type initial three-dimensional data matrix and perform time stamp and spatial coordinate calibration on it. The dimension of the multi-vehicle type initial data matrix is , is the total number of vehicle types, is the number of key power feature dimensions, is the total number of sampling times; S12: Locate the null value positions in the multi-vehicle type initial three-dimensional data matrix , obtain the values of its adjacent valid sampling points, and use at least one interpolation algorithm among linear interpolation, Newton interpolation, or cubic spline interpolation to perform interpolation compensation on the null value positions, and overwrite the interpolated data into the multi-vehicle type initial three-dimensional data matrix to generate a three-dimensional data matrix with complete records ; S13: Construct an outlier evaluation function , by accumulating the deviation values of each vehicle type in each feature dimension, and then weighting to measure the importance of different vehicle types. The expression formula of the outlier evaluation function is as follows:

[0008] Among them, represents the vehicle model in the feature dimension at time the original observation value; represents the mean value of the feature dimension in all vehicle model data, represents the vehicle model 's weight coefficient, traversing time , if exceeds the threshold , then remove the outliers of the data related to that time, and form the final three-dimensional data matrix ; S14: According to the vehicle type, add a unified encoding to the sampling data in the final three-dimensional data matrix , specifically: represents an electric vehicle when represents a fuel vehicle when represents a public transportation vehicle when; perform indexing processing on key dimensions such as power output, acceleration, vehicle start and stop times, and historical trajectories of different vehicle models; S15: Construct a multi-vehicle collaborative optimization scaling function , for perform normalization processing to obtain the matrix , the specific calculation formula is as follows:

[0009] Among them, and respectively represent the lower and upper bounds of the quantiles where the feature dimension is located in the overall data, represents the vehicle model and the feature dimension 's adaptive scaling coefficient, by setting different multi-vehicle collaborative optimization scaling functions , in the same quantile interval , perform differential balancing on the power outputs of different vehicle models; Expand in the time index dimension to obtain the multi-vehicle model feature vector sequence , among them, .

[0010] As an embodiment of the present application, the construction of the multi-vehicle model feature embedding and high-order coupling network layer based on the multi-vehicle model feature vector sequence in step S2 specifically includes: S211: Based on the embedding matrix and the learnable non-linear activation function , the multi-vehicle feature vector sequence is transformed into , and the specific calculation formula is as follows:

[0011] where represents the new multi-vehicle feature vector sequence; represents the transpose of ; represents the bias; S212: Construct a coupling matrix , and transform into for modeling the dynamic interaction effects between different vehicle models. The specific expression formula is as follows:

[0012] S213: Perform element-wise to the power and perform mapping accumulation in coordination with a correction function to additionally capture high-dimensional coupling in the continuous domain. The specific expression formula is as follows:

[0013] where is the multi-branch aggregation output, represents the number of multi-branches, represents the highest order of a set of polynomial expansions corresponding to each branch ; represents the corresponding element or matrix column mapping multiplication, and the integral term is used to dynamically correct the non-linear interaction of multi-vehicle features in different sections; S214: Construct a gating matrix , and fuse it with through element-wise multiplication to obtain the final input matrix . The specific calculation formula is as follows: .

[0014] As an embodiment of the present application, in step S2, a hierarchical skeleton network including a basic dynamics layer and a perturbation behavior layer is further set up to double-capture conventional acceleration and deceleration and short-term strong changes, and finally the outputs of the two levels are fused to obtain a preliminary prediction result, which specifically includes: S221: The final input matrix Input to the basic dynamics layer, and regress the dynamic behavior during vehicle driving, including regular acceleration and deceleration as well as relatively stable multiplicative factors during urban uniform driving, to achieve a preliminary estimate of the speed and acceleration indicators of the vehicle under normal operating conditions. The specific calculation formula is as follows:

[0015] Where, , represents the output of the basic dynamics layer; is the highest polynomial order of the model; is element-wise power; is the weight; is the bias; , is the cross-model coupling term, representing the symbiotic influence of different models under the same traffic flow, represents the all-ones vector; Based on the two-norm to measure the output error of the basic dynamics layer, the calculation formula is as follows:

[0016] Where, is the historical observation; S222: Input the final input matrix into the perturbation behavior layer, perform a non-linear mapping on the final input matrix to generate the intermediate activation , and the calculation formula is as follows:

[0017] Where, is the weight matrix, is the bias; Construct the loss function , and the calculation formula is as follows:

[0018] Where, , represents the output of the perturbation behavior layer; represents the target observation of the perturbation behavior layer; Take the partial derivative of the weight matrix through gradient update, and the calculation formula is as follows:

[0019] Where, , represents the gain matrix, which is used to adaptively amplify the gradient update of the perturbation correction layer and strengthen the response under bus stop or fast charging mode; S223: Perform additive coupling on the output of the basic dynamics layer and the output of the perturbation behavior layer to obtain the comprehensive prediction of multiple vehicle models at time , and the specific calculation formula is as follows:

[0020] where, , represents the preliminary prediction result of the total output after fusion.

[0021] As an embodiment of the present application, in step S3, merging the preliminary prediction result with the vehicle prior features to establish a multi-vehicle model coupling objective function to form an initial solution set specifically includes: S311: Merge the preliminary prediction result with the vehicle prior features to establish a multi-vehicle model coupling objective function , combine the basic dynamics cost and the scenario penalty term to form an initial solution set , which is used for global optimization of the pure electric vehicle set, fuel vehicle set, and public transportation vehicle set, and the specific formula expression is as follows:

[0022] where, represents the trajectory prediction state vector of each vehicle at the discrete time ; represents the basic dynamics evaluation of vehicle , represents the respective scenario-based penalty functions of vehicle , and the overall performance of multiple vehicle models in the time domain is characterized by cumulative integration.

[0023] As an embodiment of the present application, in step S3, constructing a vehicle type scenario-based penalty and dynamic radius mechanism, and using an improved bitterling algorithm for parallel optimization to obtain the global optimal multi-vehicle model trajectory specifically includes: S321: For the pure electric vehicle subset , construct a second-order acceleration penalty that reflects the steep change during high-speed starting, and superimpose the first derivative and the second derivative in the discrete iteration dimension to further capture the steep change error of the subtle high-speed starting. The specific calculation formula is as follows:

[0024] where, is the second-order acceleration penalty, is the iteration round of the bitterling algorithm; represents the acceleration of vehicle at the discrete iteration step , and is the positive weight coefficient. The first term characterizes the first-order severity of acceleration through discrete summation, and the second term measures the second-order mutation degree of vehicle acceleration through integration; S322: For the fuel vehicle subset In the fuel vehicle emission penalty Add coasting and emission accumulation amounts, so that solutions with excessive emission indicators or too long coasting are under elimination pressure during iterative update; S323: For the public transport vehicle subset Set the dynamic aggregation radius , and reduce it when starting and stopping frequently , so that bitterling individuals gather and search in a local area to capture sudden fluctuations; when it is relatively stable, it is relatively relaxed , to maintain the globality of multi-peak optimization; S324: Use the improved bitterling algorithm for the pure electric vehicle subset , the fuel vehicle subset and the public transport vehicle subset Embed the second-order acceleration penalty , the emission penalty and the dynamic aggregation radius respectively, and complete the iterative correction and search expansion of the trajectory parameters. The specific calculation formula is as follows:

[0025]

[0026] Among them, represents the value after updating the th vehicle parameter of the th individual in the th iteration of the bitterling algorithm; represents the multi-vehicle type solution represented by each individual in each iteration of the bitterling algorithm; is the step size decay function, represents the partial derivative of the objective function, represents the random search gain; S325: Through parallel iteration, when the maximum number of rounds is reached or the convergence threshold is met, select the solution that minimizes the multi-vehicle type coupling objective function from all individuals as the global optimal trajectory solution .

[0027] As an embodiment of the present application, the hierarchical decomposition and construction of the basic dynamics of the global optimal trajectory in step S4 specifically include: S411: Project the trajectory information in the global optimal trajectory solution onto the basic dynamics layer and the perturbation correction layer split by post-feedback respectively, and form two sets of interrelated model inputs through feedback. The basic dynamics layer is used to characterize the normal evolution of the overall traffic flow; the perturbation correction layer provides additional adjustments for extreme scenarios such as temporary bus stops or frequent speed changes. S412: The set of game participants in the basic dynamics layer is , set a residual term to evaluate the deviation between the basic layer and the actual working conditions, and regard each vehicle as a game participant, making them compete to reduce the contribution to the overall traffic deviation. Construct a multi-objective function within the time domain of to describe the above process. The specific formula is as follows:

[0028] where, represents the corresponding true observation; represents the weighted norm metric, is the weight matrix; is the game penalty function.

[0029] As an embodiment of the present application, in step S4, a residual game mechanism is constructed and the perturbation correction layer is iteratively updated. Merging the residuals of the two layers to obtain the hierarchical game residual specifically includes: S421: In the perturbation correction layer, construct hierarchical constraints to reflect the more prominent position of public transport vehicles in the game with other vehicle types. The specific formula is as follows:

[0030]

[0031] where, represents the additional perturbation correction vector of the th bus at time ; represents the secondary residual for bus perturbation; represents the perturbation correction layer penalty coefficient; is used to measure the amplitude of the correction strategy; represents the true working condition observation value of public transport vehicles; represents the time-varying trajectory of public transport vehicles; S422: When the basic dynamics layer and the perturbation correction layer complete their respective residual games, merge the residual outputs of all vehicles, and synchronously superimpose the additional perturbation correction vector of public transport vehicles onto the corresponding residual terms to generate the hierarchical game residual and the perturbation correction result of the bus set. The specific formula is as follows:

[0032]

[0033] Among them, the hierarchical game residual , the th row of which corresponds to the residual of the vehicle at each moment; represents the total residual term after merging.

[0034] As an embodiment of the present application, the coupling of the hierarchical game residual with the search process of the improved bitterling algorithm in step S5 specifically includes: S511: Establish an index table for binding the global optimal trajectory solution and the hierarchical game residual , and construct a mapping table between the vehicle and the moment . The specific formula is expressed as follows:

[0035] S512: Add additional hierarchical residual correction to the multi-vehicle coupling objective function to form a new fitness function . The fitness calculation formula on the time interval is expressed as follows:

[0036] Among them, is the trajectory candidate solution of the vehicle at the moment , represents the true observation of the vehicle ; is the local perturbation of the public transport vehicle; is the weighting matrix, represents the weighted norm squared; is the residual coupling coefficient; is the bus perturbation coupling coefficient; is the continuous integration interval for quantifying the residual additional loss under high-frequency fluctuations; represents the penalty term for evaluating the transient behavior of the vehicle within this continuous interval.

[0037] As an embodiment of the present application, the step S5 of synchronously fine-tuning the multi-vehicle trajectory parameters and residual information and outputting the optimal multi-vehicle prediction trajectory specifically includes: S521: At the initial iteration, the index table in and The mapping relationship is placed within the individual structure of Rhodeus ocellatus, and information is added to public transportation vehicles, obtaining: Information, obtaining:

[0038] ; In each round of iteration to , calculate , by performing fitness comparison on each individual in Rhodeus ocellatus, select several neighborhood leading individuals with better performance, and map their search results to the next swimming update stage. The swimming update is specifically expressed as:

[0039] Among them, is the step size factor, is the residual coupling step size, and respectively represent the residual and bus disturbance in the generation, represents the gradient of the objective function with respect to the solution vector at in the iteration; S522: Use the swimming update to calculate the latest residual of the vehicle . The specific calculation formula is as follows:

[0040] Among them, is the residual calibration coefficient. By performing quadratic iterative accumulation on the deviation between and , the hierarchical game residual information is corrected in a timely manner, enabling subsequent iterations to achieve a calibration closed-loop within the shortest range; S523: By continuously repeating the above steps until the pre-set convergence condition is met or the maximum number of iterations is reached, obtain the final trajectory of the vehicle set at different times and the corresponding residual .

[0041] The beneficial effects of the present invention are: (1)The present invention incorporates pure electric vehicles, fuel vehicles, and public transportation vehicles into the same analysis framework, and through the methods of differential processing and mutual coupling, realizes the collaborative prediction and subsequent optimization of the trajectories of multiple types of vehicles in the urban traffic system. In the data processing stage, multi-dimensional information such as the power output, acceleration, vehicle position, and start / stop time of different vehicle types is integrated. Through time alignment and spatial calibration, a feature matrix that can simultaneously reflect frequent start / stop and conventional smooth driving is formed; in this matrix, the numerical distributions of different vehicle types are significantly different. During the normalization process, a multi-vehicle collaborative scaling function is introduced to align the characteristic quantities of pure electric vehicles, fuel vehicles, and public transportation vehicles within the same quantile interval, thereby laying a consistent data foundation for subsequent modeling.

[0042] (2)The present invention predicts vehicle trajectories by adopting a hierarchical skeleton network in the modeling process. The basic dynamics layer is oriented towards common acceleration / deceleration and relatively smooth driving conditions, and captures the mainstream evolution of the vehicle speed and energy consumption through multiple power expansions and cross-vehicle coupling; comprehensively considering special disturbances such as the high-frequency stops of public transportation vehicles, a disturbance behavior layer is superimposed on the original network framework, which can actively correct for short-term or instantaneous large gradient changes, thus forming a two-layer parallel prediction structure with the basic dynamics layer. The hierarchical structure combines the characteristics of different vehicle types at the input end and generates prediction results for local mutations and overall trends simultaneously at the output end.

[0043] (3)The present invention introduces an improved bitterling algorithm in the post-processing stage, enabling the respective trajectory parameters of pure electric vehicles, fuel vehicles, and public transportation vehicles to be updated in parallel during the population search. After forming a preliminary optimal solution for the trajectory prediction results, the residual information is corrected again through the idea of hierarchical game, highlighting the disturbance impact of the short-cycle start / stop of public transportation vehicles on other vehicles, and constructing a "multi-vehicle unified modeling - hierarchical prediction - global optimization - dynamic residual correction" process, covering the key links from data preparation to online search and multi-scenario collaboration. Through the organic integration of different technical levels, the dynamic trajectory prediction and parameter optimization of vehicle groups in complex road environments are finally completed to meet the needs of the continuously evolving traffic system. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is the overall flow chart of the multi-vehicle trajectory collaborative prediction method based on bitterling optimization and hierarchical game provided in the embodiment of the present invention; Figure 2 It is the flow chart of obtaining preliminary prediction results of the multi-vehicle trajectory collaborative prediction method based on bitterling optimization and hierarchical game provided in the embodiment of the present invention; Figure 3 It is the flow chart of obtaining the globally optimal multi-vehicle trajectory of the multi-vehicle trajectory collaborative prediction method based on bitterling optimization and hierarchical game provided in the embodiment of the present invention; Figure 4 This is the flowchart for generating the hierarchical game residual matrix of the multi-vehicle trajectory collaborative prediction method based on bitterling-optimized hierarchical game provided in the embodiments of the present invention. Specific embodiments

[0045] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0046] It should be noted that all directional indications (such as up, down, left, right, front, back...) in the embodiments of the present invention are only used to explain the relative position relationship and movement conditions between components in a specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications will also change accordingly.

[0047] In the present invention, unless otherwise clearly defined and limited, terms such as "connection" and "fixation" should be understood in a broad sense. For example, "fixation" can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components or the interaction relationship between two components, unless otherwise clearly limited. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0048] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the meaning of "and / or" appearing throughout the text includes three parallel solutions. Taking "A and / or B" as an example, it includes solution A, solution B, or the solution where A and B are satisfied simultaneously. In addition, the technical solutions between various embodiments can be combined with each other, but it must be based on the fact that those of ordinary skill in the art can implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0049] Referring to Figures 1 to 4 , the first aspect of the present invention provides a multi-vehicle trajectory collaborative prediction method based on bitterling-optimized hierarchical game, and the method includes the following steps: S1: Construct an initial three-dimensional data matrix for multiple vehicle models, perform interpolation compensation on the null value records in the initial three-dimensional data matrix of multiple vehicle models, eliminate the outliers exceeding the threshold, and perform unified coding and indexing for different vehicle models; perform differential normalization processing on the feature dimension based on the multi-vehicle collaborative optimization scaling function to obtain a sequence of multi-vehicle model feature vectors; S2: Based on the sequence of multi-vehicle model feature vectors, perform multi-vehicle model feature embedding and construct a high-order coupling network layer, then construct a hierarchical skeleton network including a basic dynamics layer and a perturbation behavior layer to capture both conventional acceleration and deceleration and short-term strong changes, and finally fuse the outputs of the two levels to obtain a preliminary prediction result; S3: Combine the preliminary prediction result with the prior vehicle features, establish a multi-vehicle model coupling objective function to form an initial solution set; construct a vehicle type scenario-based penalty and dynamic radius mechanism, and use an improved bitterling algorithm for parallel optimization to obtain the global optimal multi-vehicle model trajectory; S4: Perform hierarchical decomposition on the global optimal trajectory and construct basic dynamics; construct a residual game mechanism and iteratively update the perturbation correction layer, and combine the residuals of the two layers to obtain hierarchical game residuals; S5: Couple the hierarchical game residuals with the search process of the improved bitterling algorithm, synchronously fine-tune the multi-vehicle model trajectory parameters and residual information, and output the optimal multi-vehicle model prediction trajectory; S6: Apply the optimal multi-vehicle model prediction trajectory to the short-term trajectory prediction deployment of autonomous vehicles to complete the comprehensive prediction and scheduling of different vehicle models.

[0050] As an embodiment of the present application, the step S1 specifically includes: S11: Obtain multi-dimensional information such as power output, acceleration, vehicle start and stop times, and historical trajectories of battery electric vehicles, fuel vehicles, and public transportation vehicles, organize the above multi-dimensional information in a time series, and construct an initial three-dimensional data matrix for multiple vehicle models and perform time stamp and spatial coordinate calibration on it to ensure that the time series and geographical reference information are in the same alignment standard among different vehicle models; the initial data matrix of multiple vehicle models has a dimension of , where is the total number of vehicle types, is the number of key power feature dimensions, Specifically, the timestamp calibration maps the original time data collected from different vehicle models to the same time reference, such as a unified second-level Unix timestamp or a local time zone benchmark, to ensure the comparability of all vehicles on the same sequence; the spatial coordinate calibration can convert the original longitude and latitude or local coordinates of each vehicle model to a unified geographic reference coordinate system, such as WGS84 / UTM, and perform interpolation and coordinate correction on discrete or missing position points to ensure that multi-vehicle models obtain a consistent reference standard in the spatial dimension.

[0051] S12: Locate the null value positions in the initial three-dimensional data matrix of the multi-vehicle models, obtain the values of their adjacent valid sampling points, and use at least one interpolation algorithm among linear interpolation, Newton interpolation, or cubic spline interpolation to perform interpolation compensation on the null value positions, and overwrite the interpolated data into the initial three-dimensional data matrix of the multi-vehicle models so that at each moment, a quantitative output is ensured for the number of key dynamic feature dimensions to generate a three-dimensional data matrix with complete records ; ; S13: Further, to eliminate outliers caused by sensor noise or extreme driving behaviors, construct an outlier evaluation function , which is established by accumulating the deviation values of each vehicle model in each feature dimension and then using a weighting method to measure the importance of different vehicle models. The expression formula of the outlier evaluation function is as follows:

[0052] Among them, represents the original observation value of vehicle model in feature dimension at time ; represents the mean value of feature dimension in all vehicle model data, represents the weight coefficient of vehicle model . Traverse time , if exceeds the threshold , and the threshold is given artificially or through expert experience, then it is considered that the outlier degree of the relevant data at this time is too high, and the outliers of the relevant data at this time are removed to form the final three-dimensional data matrix ; S14: Add a unified encoding to the sampling data in the final three-dimensional data matrix according to the vehicle type, specifically: represents a pure electric vehicle when represents a fuel vehicle when At the time, it represents a public transportation vehicle; index the key dimensions of power output, acceleration, vehicle start and stop times, and historical trajectories for different vehicle models to ensure that different vehicle models and different features can be indexed and queried within the same matrix structure; Further, the final three-dimensional data matrix At the time Rearrange in the order of, so that the changes of all vehicle models between and are consistent on the two-dimensional coordinate where, is used as a discrete time index, that is, the th sampling moment or time step, used to identify the data positions at different time sequences; when the final three-dimensional data matrix At the time and When rearranging between, represents the index numbers of the data at the and the th discrete moments.

[0053] S15: For the differences in maximum power output, acceleration distribution, and frequent start and stop among different vehicle models, construct a multi-vehicle collaborative optimization scaling function , for Perform normalization processing to obtain the matrix , and the specific calculation formula is as follows:

[0054] where, and respectively represent the lower and upper bounds of the quantiles of the feature dimension in the overall data, represents the vehicle model and the feature dimension 's adaptive scaling coefficient. By setting different multi-vehicle collaborative optimization scaling functions , within the same quantile interval , it is possible to differentially balance the power outputs of different vehicle models, and extreme values or frequent start and stop nodes can be included in a unified numerical range; where, refers to the final cleaned data obtained after interpolation compensation, outlier removal, and unified alignment. Compared with , is the three-dimensional data after all links of data preprocessing are completed, and is the benchmark input for subsequent normalization processing or model training.

[0055] Unfold in the time index dimension to obtain a multi-vehicle model feature vector sequence , where , specifically, the multi - vehicle type feature vector sequence covers power output, acceleration, vehicle position, and start - stop event information, and is uniformly mapped on the numerical scale.

[0056] Specifically, in the prior art, when facing parallel data of multiple vehicle types, fixed minimum or maximum normalization is mostly used, which does not flexibly adjust the dynamic characteristics differences between vehicle types, resulting in large differences in the numerical scale after normalization; in this application, a multi - vehicle collaborative optimization scaling function is constructed , overcoming the limitations of insufficient start - stop frequency and acceleration jump of different vehicle types, ensuring that the mapped values of different vehicle types on the same feature dimension can be compared equally, and then realizing the stability of the network training and prediction process in the multi - vehicle type mixed scenario. It also allows users to customize the importance of some vehicle types on different driving roads.

[0057] In step S2 of this application, based on the multi - vehicle type feature vector sequence output in step S1, a hierarchical skeleton network is proposed to preliminarily fit the vehicle dynamics evolution; the hierarchical skeleton network generally includes two main functional layers: the basic dynamics layer and the perturbation behavior layer, which are respectively aimed at double - capturing conventional acceleration and deceleration and short - time local strong changes; through the fusion output of the multi - layer structure, collaborative prediction of multiple vehicle types (such as pure electric vehicles, fuel vehicles, public transportation vehicles) in different scenarios can be realized.

[0058] As an embodiment of this application, in step S2, based on the multi - vehicle type feature vector sequence, the construction of multi - vehicle type feature embedding and high - order coupling network layer specifically includes: S211: Based on the embedding matrix and the learnable non - linear activation function , the multi - vehicle type feature vector sequence is transformed into a new multi - vehicle type feature vector sequence , and the specific calculation formula is as follows:

[0059] where represents the new multi - vehicle type feature vector sequence; represents the transpose of ; represents the bias; S212: Further considering the mutual perturbation problem between vehicle types, a coupling matrix is constructed, and is transformed into for modeling the dynamic interaction effects between different vehicle types. The specific expression formula is as follows:

[0060] Among them, represents the newly encoded multi-vehicle feature vector sequence after considering the coupling or interference relationship between different vehicle models, which is further obtained by matrix mapping with the coupling matrix to obtain the latest multi-vehicle feature vector sequence representation, used to characterize the dynamic interaction effects between vehicle models, and reflects the new features under the coupling action of vehicle models by

[0061] performing a linear transformation. Based on the above, to measure the energy linkage between the initial pure electric vehicle and the fuel vehicle, and the additional influence of the bus on the surrounding traffic flow, it is equivalent to adding an additional fully connected layer, and realizing multi-vehicle feature coupling through the linear transformation of the fully connected type.

[0062] S213: For element by element to the power of perform mapping accumulation and cooperate with the correction function to additionally capture high-dimensional coupling in the continuous domain. The specific expression formula is as follows:

[0063] Among them, is the multi-branch aggregation output, represents the number of multi-branches, represents each branch corresponding to the highest order of a set of polynomial expansions, represents the corresponding element or matrix column mapping multiplication, and the integral term is used to dynamically correct the non-linear interaction of multi-vehicle features under different sections, taking into account the frequent start and stop of public transportation and the complex mode of large acceleration of pure electric vehicles; S214: Further, to adapt to the time-varying scenario, construct a gating matrix , and through element-by-element multiplication with to fuse and obtain the final input matrix . The specific calculation formula is as follows:

[0064] Among them, represents the column dimension of the matrix; that is, when there are special situations such as local congestion or extreme climate, can adaptively amplify or suppress some vehicle model channels, and then dynamically regulate the significance of multi-vehicle feature components in terms of time; will then serve as the high-dimensional input to the basic dynamics layer and the perturbation behavior layer in the next step to complete the compatibility and difference capture of the multi-vehicle situation.

[0065] As an embodiment of the present application, in step S2, a hierarchical skeleton network including a basic dynamics layer and a perturbation behavior layer is constructed to capture both conventional acceleration and deceleration and short-term strong changes, and finally the outputs of the two levels are fused to obtain a preliminary prediction result, which specifically includes: S221: Input the final input matrix into the basic dynamics layer to regress the dynamic behaviors in most vehicle driving periods, including conventional acceleration and deceleration as well as relatively stable multiplicative factors during uniform driving in urban areas, and realize the preliminary estimation of the speed and acceleration indexes of the vehicle under normal working conditions. The specific calculation formula is as follows:

[0066] Where , represents the output of the basic dynamics layer; is the highest polynomial order of the model; is to the power of the element-by-element times; is the weight; by mapping and accumulating and the corresponding , capture the stable acceleration and deceleration characteristics at different orders; is the bias; , is the cross-model coupling term, representing the symbiotic influence of different models under the same traffic flow; represents the all-1 vector, and a constant offset is introduced based on the all-1 vector ; Based on the second norm to measure the output error of the basic dynamics layer, the calculation formula is as follows:

[0067] Where is the historical observation or annotation; Directly based on stochastic gradient descent, jointly update , and so that the basic dynamics layer can track the speed changes and acceleration trends in the conventional interval and form a consistent stable regression among multiple models.

[0068] S222: Input the final input matrix into the perturbation behavior layer to perform a non-linear mapping on the final input matrix to generate an intermediate activation , and the calculation formula is as follows:

[0069] Where is the weight matrix, is the bias; Construct a loss function , the calculation formula is as follows:

[0070] Among them, , represents the output of the perturbation behavior layer; represents the target observation of the perturbation behavior layer; Perform additional enhancement on the features in the high-frequency or instantaneous change interval to construct a gain matrix , and take the partial derivative of the weight matrix through gradient update. The calculation formula is as follows:

[0071] Among them, , represents the gain matrix, which is used to adaptively amplify the gradient update of the perturbation correction layer and strengthen the response under the bus stop or fast charging mode (instantaneous acceleration of pure electric vehicles); specifically, when a large gradient error is induced for a certain vehicle model, the gain matrix can significantly amplify it to ensure more fine-grained capture of instantaneous phenomena such as rapid acceleration of buses or pure electric vehicles.

[0072] This step is mainly used to capture frequent stops and local mutations in public transportation, and specifically model and correct the high-frequency start-stop of public transportation vehicles or the sharp increase in instantaneous acceleration of pure electric vehicles to make up for the deficiencies of the basic layer in local large gradient regions.

[0073] S223: Perform additive coupling on the output of the basic dynamics layer and the output of the perturbation behavior layer to obtain the comprehensive prediction of multiple vehicle models at time . The specific calculation formula is as follows:

[0074] Among them, , represents the preliminary prediction result of the total output after fusion.

[0075] Take as the overall loss, and synchronously train all parameters (including , , , , , etc., related parameters that need to be iteratively trained); that is, in the inference stage, for the given multi-vehicle model features after feature embedding and coupling to obtain , and then send them into the basic dynamics layer and the perturbation behavior layer respectively, and finally additively fuse to obtain the preliminary prediction value . Through this step, preliminary collaborative prediction can be carried out for pure electric vehicles, fuel vehicles, and public transportation vehicles in a unified framework, taking into account the needs of smooth driving and local mutations.

[0076] As an embodiment of the present application, in step S3, merging the preliminary prediction result with the vehicle prior features to establish a multi-vehicle type coupling objective function to form an initial solution set specifically includes: S311: Merge the preliminary prediction result with the vehicle prior features. The vehicle prior features refer to the basic physical parameters or historical statistical information that the vehicle already has before entering the model and are used to assist in optimization or evaluation; establish a multi-vehicle type coupling objective function , combine the basic dynamics cost with the scenario penalty term to form an initial solution set , for global optimization of the pure electric vehicle set, fuel vehicle set, and public transportation vehicle set, and the specific formula expression is as follows:

[0077] Wherein, represents the trajectory prediction state vector of each vehicle at discrete time ; represents the basic dynamics evaluation of vehicle , represents the respective scenario-based penalty functions of vehicle , and the overall performance of multi-vehicle types in the time domain is characterized by cumulative integration.

[0078] Specifically, for a population size of , each population individual corresponds to a set of multi-vehicle type trajectory parameters { , , ,…, }, enabling the bitterling algorithm to make full use of the difference information of various types of vehicles during update.

[0079] Traditional multi-vehicle type optimization often uses single penalty or unified threshold, which cannot accurately reflect the differences in the high-speed acceleration of pure electric vehicles and the frequent starting and stopping of public transportation vehicles, and is prone to local convergence or incomplete optimization; the present invention can couple the key features of different vehicle types into a unified index system by introducing graded penalty and performing integral accumulation, which not only retains the integrity of global search but also provides an accurate tuning space for subsequent processing.

[0080] As an embodiment of the present application, in step S3, constructing a vehicle type scenario-based penalty and dynamic radius mechanism, and using an improved bitterling algorithm for parallel optimization to obtain the global optimal multi-vehicle type trajectory specifically includes: S321: Construct a second-order acceleration penalty reflecting the steep change in high-speed starting for the pure electric vehicle subset , and superimpose the first derivative and the second derivative in the discrete iteration dimension, further capture the steep change error in the subtle high-speed start, and the specific calculation formula is as follows:

[0081] Among them, is the second-order acceleration penalty, is the iteration round of the bitterling fish algorithm; represents the vehicle at the discrete iteration step under the acceleration, and are the forward weight coefficients, represents the first-order intensity of the acceleration characterized by discrete summation, represents the second-order mutation degree of the vehicle acceleration measured by integration; S322: Add the coasting and emission accumulation amount to the fuel vehicle subset in the fuel emission penalty , and the is an additional penalty for the solution of the fuel vehicle and an emission adjustment factor, so that the solutions with excessive emission indicators or long-term coasting are under elimination pressure during iterative update; S323: Set the dynamic aggregation radius for the public transport vehicle subset , and reduce when the public transport vehicles start and stop frequently, so that the bitterling fish individuals gather and search in the local area to capture sudden fluctuations; when it is relatively stable, is relatively relaxed to maintain the global nature of multi-peak optimization; Specifically, by introducing the fuel emission penalty , the dynamic aggregation radius re-adjust the initial solution set , and perform local or global update on the initial solution set to refine the solution set according to the new constraints or adjustment factors before entering the subsequent global optimization, so as to better meet the actual operation requirements.

[0082] S324: Use the improved bitterling fish algorithm to embed the second-order acceleration penalty , the fuel vehicle subset and the public transport vehicle subset into the second-order acceleration penalty , the emission penalty and the dynamic aggregation radius respectively, and complete the iterative correction and search expansion of the trajectory parameters. The specific calculation formula is as follows:

[0083]

[0084] Among them, represents the value after the Rhodeus algorithm updates the th vehicle parameter of the th individual in the th iteration; represents the multi-vehicle type solution represented by each individual in each iteration of the Rhodeus algorithm; is the step size decay function, represents the partial derivative of the objective function, represents the random search gain to prevent falling into local extrema; S325: Through parallel iteration, when the maximum number of rounds is reached or the convergence threshold is met, select the solution that minimizes the multi-vehicle type coupling objective function from all individuals as the global optimal trajectory solution , and submit the global optimal trajectory solution to the next step for further refinement using the game residual strategy at a higher level.

[0085] As an embodiment of the present application, the hierarchical decomposition and construction of the basic dynamics for the global optimal trajectory in step S4 specifically include: S411: Project the trajectory information in the global optimal trajectory solution onto the basic dynamics layer and the perturbation correction layer split through post-feedback respectively, and further feedback to form two sets of mutually related model inputs. The basic dynamics layer is used to characterize the normal evolution of the overall traffic flow; the perturbation correction layer provides additional adjustments for extreme scenarios such as temporary bus stops or frequent speed changes.

[0086] Specifically, based on the multi-layer structure of the initial prediction network constructed in step 2, the idea of game residual correction is further carried out. In this step, instead of further training and searching for the same two-layer network, the already obtained global trajectory is post-feedback split into a basic dynamics layer and a perturbation correction layer for residual game and generating new residual outputs.

[0087] Specifically, step 2 performs preliminary network layering facing the model input end to distinguish different types of working conditions in the early prediction; step 4 faces the output end and performs targeted game-based residual correction on the already obtained optimal trajectory to further approximate the real working conditions in the later stage; through such a structural design, the accuracy, robustness, and applicability of multi-vehicle type trajectory prediction and optimization can be improved.

[0088] S412: The set of game participants in the basic dynamics layer is , a residual term is set to evaluate the deviation between the basic layer and the actual working conditions, and each vehicle is regarded as a game participant, making them compete to reduce the contribution to the overall traffic deviation. In Construct a multi-objective function in the time domain to characterize the above process. The specific formula is as follows:

[0089] Wherein, represents the trajectory prediction state vector of each vehicle at discrete time ; represents the corresponding true observation; represents the weighted norm metric, is the weight matrix; is the game penalty function. If the residuals of some vehicles accumulate significantly at time , additional constraints are imposed on them.

[0090] As an embodiment of the present application, in step S4, a residual game mechanism is constructed and the perturbation correction layer is iteratively updated. Merging the residuals of the two layers to obtain the hierarchical game residuals specifically includes: S421: In the perturbation correction layer, construct hierarchical constraints to reflect the more prominent position of public transport vehicles in the game with other vehicle types. The specific formula is as follows:

[0091]

[0092] Wherein, represents the additional perturbation correction vector of the th bus at time ; represents the secondary residual for bus perturbation; represents the perturbation correction layer penalty coefficient; is used to measure the amplitude of the correction strategy; represents the true operating condition observation value of the public transport vehicle; represents the time-varying trajectory of the public transport vehicle; the optimal solution tends to moderately control the energy consumption of the correction while ensuring is small, and maintain a low additional interference to the base layer.

[0093] Specifically, in this step, for the subset of buses, the remaining vehicles are mainly based on the base layer. The time-varying trajectory of the buses is further analyzed, and variable weight penalties are used to strengthen the rapid correction of frequent stops, acceleration and deceleration; since the perturbation layer is coupled with the base layer in a local time period, a residual competition mechanism is established: when the local perturbation is too large, the influence of public transport vehicles among surrounding vehicles is preferentially reduced; when the perturbation is in small fluctuations, it is encouraged to approach the base operating condition faster.

[0094] In the conventional multi - vehicle model method, it is easy to overlook the cascading effect of sudden bus stops on other vehicle types or to cause local delays by centralized correction. In this application, the bus disturbance layer is separately extracted through hierarchical game and synchronized with the basic layer for game, which can not only ensure the disturbance correction speed of bus vehicles, but also reduce the impact on the overall traffic flow.

[0095] S422: When the basic dynamics layer and the disturbance correction layer complete their respective residual games, the residual outputs of all vehicles are merged, and the additional disturbance correction vectors of the public transport vehicle subset are synchronously superimposed on the corresponding residual terms. The fusion process takes into account the different performance of public transport vehicles, pure electric vehicles and fuel vehicles at different times, forming a complete correction under the condition of multi - vehicle mixed traffic; finally, a hierarchical game residual matrix and the disturbance correction results of the bus set are generated. The specific formula is as follows:

[0096]

[0097] Among them, the hierarchical game residual matrix covers the final hierarchical game residual information of all vehicles. Its th row corresponds to the residual of vehicle at each moment; represents the combined total residual term.

[0098] Specifically, in this application, by introducing a hierarchical game and a residual competition strategy between the basic dynamics layer and the disturbance correction layer, it is ensured that the basic dynamics correction of pure electric vehicles and fuel vehicles and the local disturbance correction of bus vehicles can be quickly and accurately learned.

[0099] As an embodiment of this application, the coupling of the hierarchical game residual and the search process of the improved bitterling algorithm in step S5 specifically includes: S511: Establish an index table , which is used to bind the global optimal trajectory solution and the hierarchical game residual matrix , and construct a mapping table for vehicle and time . The specific formula is as follows:

[0100] Add in parallel to the index row corresponding to the bus vehicle; S512: Add additional hierarchical residual correction to the multi - vehicle coupling objective function to form a new fitness function , in the time interval the fitness calculation formula is expressed as follows:

[0101] where, is the trajectory candidate solution of the vehicle at the moment ; represents the true observation of the vehicle ; is the local perturbation of the public transport vehicle (for ordinary vehicles, let ), is the weighting matrix, represents the weighted norm square; is the residual coupling coefficient; is the bus perturbation coupling coefficient; is the continuous integration interval, which is used to quantify the residual additional loss under high-frequency fluctuations; represents the penalty term for evaluating the transient behavior of the vehicle in this continuous interval.

[0102] As an embodiment of the present application, the step S5 of synchronously fine-tuning the multi-vehicle trajectory parameters and residual information and outputting the optimal multi-vehicle prediction trajectory specifically includes: S521: At the initial iteration, the mapping relationship between in the index table and is placed in the bitterling individual structure, and information is added to the public transport vehicle to obtain:

[0103] ; where, represents the initial trajectory solution of the th vehicle at the moment in the initial iteration round , which is directly taken from ; represents the initial dynamic radius of the th vehicle at the moment in the initial iteration round , which is directly taken from ; In each iteration to , calculate the objective function corresponding to the solution obtained in the , by performing fitness comparison on each individual in the bitterling fish, several neighborhood leading individuals with better performance are selected, and their search results are mapped to the next swimming update stage. The swimming update is specifically expressed as:

[0104] Among them, is the step size factor, is the residual coupling step size, and respectively represent the residual and bus disturbance in the generation, represents that in the round of iteration, the objective function with respect to the solution vector at is the gradient, which is used to guide how the bitterling fish individuals swim or update; S522: Use the swimming update to calculate the latest residual of the vehicle . The specific calculation formula is as follows:

[0105] Among them, is the residual calibration coefficient. By performing quadratic iterative accumulation on the deviation between and , the hierarchical game residual information is corrected in a timely manner, enabling subsequent iterations to achieve a correction closed-loop within the shortest range; S523: By continuously repeating the above steps until the pre-set convergence condition is met or the maximum number of iterations is reached, the final trajectories of the vehicle set at different times and the corresponding residuals are obtained.

[0106] Output the finally converged and the finally converged dynamic radius . Save the final local disturbance of the bus vehicle together, which is used for the trajectory scheduling of mixed operation of multiple vehicle types or subsequent energy consumption management.

[0107] In the present invention, by introducing the residual and the local disturbance of the public transportation vehicle in real time during the swimming process of the bitterling fish, the dynamic compatibility between the public transportation vehicle and the pure electric vehicle and the fuel vehicle is realized, overcoming the limitation that the existing solution cannot accurately adapt to the temporary stop or frequent speed change of the public transportation vehicle.

[0108] Finally, the output result of step S5 is deployed in the trajectory prediction of autonomous vehicles, that is, after the bitterling optimization and hierarchical game residual coupling in step S5, the final driving trajectories of all vehicles at the time to and the corresponding residual correction amounts within the range are obtained.

[0109] The predicted trajectories of electric vehicles, fuel vehicles, and bus vehicles are stored uniformly, obtaining a set of future motion states of multiple vehicle models that are closer to the actual working conditions, meeting the high-precision and high-reliability requirements of the autonomous driving system for short-term trajectory prediction.

[0110] The above description is only some preferred embodiments of the present disclosure and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A multi-vehicle trajectory collaborative prediction method based on the optimization of the hierarchical game of Rhodeus fish, characterized in that The method includes the following steps: S1: Construct an initial three-dimensional data matrix for multiple vehicle models, perform interpolation compensation on the null value records in the initial three-dimensional data matrix for multiple vehicle models, eliminate the outliers exceeding the threshold, and perform unified coding and indexing for different vehicle models; perform differential normalization processing on the feature dimension based on the multi-vehicle collaborative optimization scaling function to obtain a multi-vehicle model feature vector sequence; S2: Based on the multi-vehicle model feature vector sequence, perform multi-vehicle model feature embedding and construction of a high-order coupling network layer, and then construct a hierarchical skeleton network including a basic dynamics layer and a perturbation behavior layer to capture both conventional acceleration and deceleration and short-term strong changes, and finally fuse the outputs of the two levels to obtain a preliminary prediction result; S3: Merge the preliminary prediction result with the vehicle prior features, establish a multi-vehicle model coupling objective function to form an initial solution set; construct a vehicle type scenario-based penalty and dynamic radius mechanism, and use an improved bitterling algorithm for parallel optimization to obtain a globally optimal multi-vehicle model trajectory; S4: Perform hierarchical decomposition on the globally optimal trajectory and construct a basic dynamics layer; construct a residual game mechanism and perform iterative update on the perturbation correction layer, and merge the residuals of the two layers to obtain a hierarchical game residual; S5: Couple the hierarchical game residual with the search process of the improved bitterling algorithm, synchronously fine-tune the multi-vehicle model trajectory parameters and residual information, and output an optimal multi-vehicle model prediction trajectory; S6: Apply the optimal multi-vehicle model prediction trajectory to the short-term trajectory prediction deployment of autonomous vehicles to complete the comprehensive prediction and scheduling of different vehicle models.

2. The multi-vehicle trajectory collaborative prediction method based on optimized hierarchical game of Rhodeus ocellatus according to claim 1, wherein The specific steps of S1 include: S11: Obtain the multi-dimensional information of power output, acceleration, vehicle start and stop times, and historical trajectory of battery electric vehicles, fuel vehicles, and public transportation vehicles. Organize the above multi-dimensional information in a time series to construct an initial three-dimensional data matrix for multiple vehicle models and calibrate its timestamp and spatial coordinates. The initial data matrix for multiple vehicle models has dimensions of , where is the total number of vehicle types, is the number of key power feature dimensions, and is the total number of sampling times; S12: Locate the null value positions in the initial three-dimensional data matrix for multiple vehicle models, obtain the values of their adjacent valid sampling points, and use at least one interpolation algorithm among linear interpolation, Newton interpolation, or cubic spline interpolation to perform interpolation compensation on the null value positions, and overwrite the interpolated data onto the initial three-dimensional data matrix for multiple vehicle models to generate a three-dimensional data matrix with complete records ; ; S13: Construct an outlier evaluation function , by accumulating the deviation values of each vehicle model in each feature dimension, and then weighing the importance of different vehicle models through weighting. The expression formula of the outlier evaluation function is as follows: Among them, represents the vehicle model at the feature dimension at the moment the original observation value; represents the feature dimension the mean value in all vehicle model data, represents the vehicle model the weight coefficient of, traverse the moment , if exceeds the threshold , then remove the outliers of the data at this moment, and form the final three-dimensional data matrix ; S14: Add unified coding to the sampling data in the final three-dimensional data matrix according to the vehicle type, specifically as follows: When [condition], it represents an electric vehicle, When [condition], it represents a fuel vehicle, When [condition], it represents a public transportation vehicle; Index the key dimensions of power output, acceleration, vehicle start and stop times, and historical trajectory for different vehicle models; ​ S15: Construct a multi-vehicle collaborative optimization scaling function , for perform normalization to obtain matrix , and the specific calculation formula is as follows: Among them, and respectively represent the lower and upper bounds of the quantiles where the feature dimension is located in the overall data. represents the vehicle model and the adaptive scaling coefficient of the feature dimension . By setting different multi-vehicle collaborative optimization scaling functions , the respective power outputs of different vehicle models are differentially balanced within the same quantile interval . Expand on the time index dimension to obtain a multi-vehicle feature vector sequence , where .

3. The multi-vehicle trajectory collaborative prediction method based on optimized hierarchical game of Rhodeus fish according to claim 2, wherein In step S2, based on the multi-vehicle model feature vector sequence, the specific steps of performing multi-vehicle model feature embedding and constructing a high-order coupling network layer include: S211: Based on the embedding matrix and a learnable non-linear activation function , transform the multi-vehicle feature vector sequence to . The specific calculation formula is as follows: Among them, represents a new multi-vehicle feature vector sequence, represents the transpose operation; represents the bias; S212: Construct a coupling matrix , and transfer to for modeling the dynamic interaction effects between different vehicle models. The specific expression formula is as follows: S213: For element by element exponentiation perform mapping accumulation, collaborate with the correction function, and additionally capture high-dimensional coupling in the continuous domain. The specific expression formula is as follows: Among them, is the multi-branch aggregated output, represents the number of multi-branches, represents each branch corresponding to the highest order of a set of polynomial expansions, represents the multiplication of corresponding element or matrix column mapping, and the integral term is used to dynamically correct the non-linear interaction of multi-vehicle features under different sections; S214: Construct a gating matrix , through element-wise multiplication with fusion to obtain the final input matrix , and the specific calculation formula is as follows: 。 4. The multi-vehicle trajectory collaborative prediction method based on the optimized hierarchical game of Rhodeus ocellatus according to claim 3, characterized in that In step S2, further setting up a hierarchical skeleton network including a basic dynamics layer and a perturbation behavior layer to capture both conventional acceleration and deceleration and short-term strong changes, and finally fusing the outputs of the two levels to obtain a preliminary prediction result specifically includes: S221: Input the final input matrix into the basic dynamics layer to regress the dynamic behavior during the vehicle driving period, including the relatively stable multiplicative factors of normal acceleration and deceleration as well as uniform driving in urban areas, and realize the preliminary estimation of the speed and acceleration indexes of the vehicle under normal working conditions. The specific calculation formula is as follows: Among them, , representing the output of the basic dynamics layer; is the highest polynomial order of the model; is 's element-wise -th power; is the weight; is the bias; , as the cross-model coupling term, represents the symbiotic influence of different vehicle models under the same traffic flow; represents the all-ones vector; Based on the two-norm to measure the output error of the basic dynamics layer, the calculation formula is as follows: Among them, is the historical observation; S222: Input the final input matrix into the perturbation behavior layer, and perform a non-linear mapping on the final input matrix to generate an intermediate activation , and the calculation formula is as follows: Among them, is the weight matrix, is the bias; Construct the loss function , and the calculation formula is as follows: Among them, represents the output of the perturbation behavior layer; represents the target observation of the perturbation behavior layer; Derive the partial derivative of the weight matrix through gradient update The calculation formula is as follows: Among them, , representing the gain matrix, is used to adaptively amplify the gradient update of the perturbation correction layer and strengthen the response under the bus stop or fast charging mode; S223: Perform additive coupling on the output of the basic dynamics layer and the output of the disturbance behavior layer to obtain the comprehensive prediction of multiple vehicle models at time , and the specific calculation formula is as follows: Among them, represents the preliminary prediction result of the total output after fusion.

5. The multi-vehicle trajectory collaborative prediction method based on optimized hierarchical game of Rhodeus ocellatus according to claim 4, characterized in that, In step S3, merging the preliminary prediction result with the vehicle prior features, and the specific steps of establishing a multi-vehicle model coupling objective function to form an initial solution set include: S311: Merge the preliminary prediction result with the vehicle prior features to establish a multi-vehicle type coupling objective function , combine the basic dynamics cost with the scenario penalty term to form an initial solution set , which is used for global optimization of the pure electric vehicle set, fuel vehicle set, and public transportation vehicle set. The specific formula is as follows: Among them, represents the trajectory prediction state vector of each vehicle at discrete moments ; represents the basic dynamic evaluation of the vehicle ; represents the respective scenario-based penalty functions of the vehicles which characterize the overall performance of multiple vehicle types in the time domain through cumulative integration.

6. The multi-vehicle trajectory collaborative prediction method based on the optimized hierarchical game of Rhodeus ocellatus according to claim 5, characterized in that, In step S3, constructing a vehicle type scenario-based penalty and dynamic radius mechanism, and the specific steps of using an improved bitterling algorithm for parallel optimization to obtain a globally optimal multi-vehicle model trajectory include: S321: For the pure electric vehicle subset Construct a second-order acceleration penalty that reflects the steep change during high-speed start-up, and superimpose the first derivative and the second derivative in the discrete iteration dimension to further capture the steep change error of the subtle high-speed start-up. The specific calculation formula is as follows: Among them, is the second-order acceleration penalty, is the iteration round of the bitterling fish algorithm; represents the vehicle at the discrete iteration step under the acceleration, and is the positive weight coefficient, characterizes the first-order severity of the acceleration through discrete summation, measures the second-order mutation degree of the vehicle acceleration through integration; S322: For the fuel vehicle subset During the fuel vehicle emission penalty Add coasting and emission accumulation amounts, so that solutions with exceeded emission indicators or overly long coasting are under elimination pressure during iterative updates; S323: For the subset of public transportation vehicles Set the dynamic aggregation radius , and reduce it when starting and stopping frequently , so that Rhodeus ocellatus individuals gather and search in a local area to capture sudden fluctuations; when it is relatively stable, it is relatively relaxed , to maintain the global nature of multi-peak optimization; S324: Use the improved Rhodeus algorithm for the pure electric vehicle subset , the fuel vehicle subset and the public transportation vehicle subset are respectively embedded with second-order acceleration penalty , emission penalty and dynamic aggregation radius , and the iterative correction and search expansion of the trajectory parameters are completed. The specific calculation formula is as follows: Among them, represents the value after the Rhodeus algorithm updates the th vehicle parameter of the th individual in the th iteration; represents the multi-vehicle solution represented by each individual in each iteration of the Rhodeus algorithm; is the step size decay function, represents the partial derivative of the objective function, represents the random search gain; S325: Through parallel iteration, when the maximum number of rounds is reached or the convergence threshold is met, select the solution that minimizes the multi-vehicle coupling objective function from all individuals as the global optimal trajectory solution .

7. The multi-vehicle trajectory collaborative prediction method based on the optimized hierarchical game of Rhodeus ocellatus according to claim 6, characterized in that, In step S4, the specific steps of performing hierarchical decomposition on the globally optimal trajectory and constructing a basic dynamics layer include: S411: Project the trajectory information in the globally optimal trajectory solution onto the basic dynamics layer and the perturbation correction layer split by post-feedback respectively. The basic dynamics layer is used to characterize the normal evolution of the overall traffic flow; the perturbation correction layer provides additional adjustment for extreme scenarios such as temporary bus stops or frequent speed changes. S412: The set of game participants in the basic dynamics layer is , and a residual term is set to evaluate the deviation between the basic layer and the actual working conditions. Each vehicle is regarded as a game player, and they are made to compete to reduce their contributions to the overall traffic deviation. A multi-objective function is constructed within the time domain of to describe the above process, and the specific formula is expressed as follows: Among them, represents the trajectory prediction state vector of each vehicle at discrete moments ; represents the corresponding true observation; represents the weighted norm metric, is the weight matrix; is the game penalty function.

8. The multi-vehicle trajectory collaborative prediction method based on the optimized hierarchical game of Rhodeus ocellatus according to claim 7, wherein In step S4, the specific steps of constructing a residual game mechanism and performing iterative update on the perturbation correction layer, and merging the residuals of the two layers to obtain a hierarchical game residual include: S421: In the perturbation correction layer, construct a hierarchical constraint to reflect the more prominent position of the set of public transportation vehicles in the game with other vehicle models. The specific formula expression is as follows: Among them, represents the additional disturbance correction vector of the th bus at time ; represents the local residual for the disturbance of public transport vehicles; represents the penalty coefficient of the disturbance correction layer; represents the magnitude used to measure the correction strategy; represents the true working condition observation value of public transport vehicles; represents the time-varying trajectory of public transport vehicles; S422: After the basic dynamics layer and the perturbation correction layer complete their respective residual games, the residual outputs of all vehicles are merged, and the additional perturbation correction vectors of public transport vehicles are synchronously superimposed on the corresponding residual terms to generate a hierarchical game residual matrix and the perturbation correction results of the bus set. The specific formula is expressed as follows: Among them, the hierarchical game residual matrix , where the -th row corresponds to the residuals of the vehicle at each moment; represents the total residual term after merging.

9. The multi-vehicle trajectory collaborative prediction method based on the optimized hierarchical game of Rhodeus ocellatus according to claim 8, wherein In step S5, the coupling of the hierarchical game residuals with the search process of the improved bitterling algorithm specifically includes: S511: Establish an index table for binding the globally optimal trajectory solution and the hierarchical game residual matrix , and in the vehicle and time Construct a mapping table, and the specific formula is expressed as follows: S512: For the multi-vehicle coupling objective function add hierarchical residual correction additionally to form a new fitness function , in the time interval the fitness calculation formula is expressed as follows: Among them, is the trajectory candidate solution of the vehicle at time . represents the true observation of the vehicle . is the local perturbation of the public transport vehicle; is the weighting matrix, which represents the weighted norm squared; is the residual coupling coefficient; is the bus perturbation coupling coefficient; is the continuous integration interval, which is used to quantify the residual additional loss under high-frequency fluctuations; represents the penalty term for evaluating the transient behavior of the vehicle within this continuous interval.

10. The multi-vehicle trajectory collaborative prediction method based on the optimized hierarchical game of Rhodeus ocellatus according to claim 9, wherein In step S5, the synchronous fine-tuning of the multi-vehicle trajectory parameters and residual information and the output of the optimal multi-vehicle prediction trajectory specifically include: S521: At the initial iteration, place the mapping relationship in the bitterling individual structure in the index table and add information to the public transportation vehicle to obtain: ​ ； In each iteration to when calculating , by performing fitness comparison on each individual in Rhodeus, several neighboring leading individuals with better performance are selected, and their search results are mapped to the next swimming update stage. The swimming update is specifically represented as: Among them, is the step size factor, is the residual coupling step size, and respectively represent the residual and bus disturbance in the th generation, represents that at the th iteration, the objective function with respect to the solution vector at is the gradient; S522: Update using random walk Calculate for the vehicle the latest residual , and the specific calculation formula is as follows: Among them, is the residual calibration coefficient. By performing quadratic iterative accumulation on the deviation between and , the residual information of the hierarchical game is corrected in a timely manner, enabling the subsequent iteration to achieve a calibration closed-loop within the shortest range; S523: By continuously repeating the above steps until the pre-set convergence condition is met or the maximum number of iterations is reached, the final trajectories of the vehicle set at different times are obtained and the corresponding residuals .

Images