Pantograph control method and device

Abstract

The invention discloses a pantograph control method and device, and relates to the field of data fusion, and the method comprises the steps: determining the weight of a prediction point in an interpolation region under each to-be-aggregated agent model according to the weight of each training sample point; determining the weight of a prediction point in the extrapolation region under each to-be-aggregated agent model according to the accuracy of each to-be-aggregated agent model; and aggregating the responses output by the plurality of agent models to be aggregated based on the weights and the output responses of the prediction points in the interpolation region and the weights and the output responses of the prediction points in the extrapolation region. And assuming that the weight of the prediction point is in direct proportion to the gradient of the weight of the training sample point, determining the weight of the prediction point in the interpolation region based on the weight of the training sample point. Since the extrapolation region is not trained, the weight is determined based on the accuracy of the proxy model. The weights are given to different agent models, so that evaluation and weight proportion adjustment among multiple models are better carried out, and the output network flow and network voltage of the pantograph are more accurate.

Description

Technical Field

[0001] The present invention relates to the field of data fusion, and particularly to a control method and device for a pantograph. Background Art

[0002] In high-speed railways or urban rail transit, the pantograph is the core device for a train to obtain electric energy from the catenary. Its dynamic performance directly affects the stability and safety of electric energy transmission, and the network current (catenary current) and network voltage (catenary voltage) are key parameters. However, the interaction between the pantograph and the catenary is affected by multiple factors (such as lift force, running speed, ambient temperature, catenary suspension height, etc.), and it is difficult for traditional single models to accurately predict the network current and network voltage under complex working conditions. Through multi-agent model integration technology, the advantages of different agent models can be dynamically fused to improve the prediction accuracy and robustness. Usually, the method for determining the weighting coefficient is average integration, that is, a global metric method is used to determine the weight coefficient of each agent model. In such a combination, the weight coefficient remains fixed in the entire design space. However, using a constant weight coefficient combination cannot fully utilize the local characteristics of each agent model, thus limiting the further improvement of the prediction accuracy, and further resulting in a large prediction error for the problems of the network current and network voltage of the pantograph. Summary of the Invention

[0003] The object of the present invention is to provide a control method and device for a pantograph, which can better evaluate between multiple models and adjust the weight ratio by assigning weights to different agent models, and the output network current and network voltage of the pantograph are more accurate.

[0004] To solve the above technical problems, the present invention provides a control method for a pantograph, including:

[0005] Determine each training sample point and each prediction point, and each of the prediction points and each of the training sample points includes at least two of the pantograph lift force, train running speed, ambient temperature, and catenary suspension height;

[0006] Determine the weights of each training sample point for training the agent models to be aggregated;

[0007] Determine the weights of the prediction points within the interpolation region under each agent model to be aggregated according to the weights of each training sample point, and the range where the training sample points are located is the interpolation region;

[0008] Determine the weights of the prediction points within the extrapolation region under each agent model to be aggregated according to the accuracy of each of the agent models to be aggregated, and the range outside the range where the training sample points are located is the extrapolation region;

[0009] Aggregate the responses output by multiple proxy models to be aggregated based on the weights of the prediction points in the interpolation region under each of the proxy models to be aggregated, the responses of the prediction points in the interpolation region output by each of the proxy models to be aggregated, the weights of the prediction points in the extrapolation region under each of the proxy models to be aggregated, and the responses of the prediction points in the extrapolation region output by each of the proxy models to be aggregated, where the response is the network current and network voltage of the pantograph;

[0010] Output the aggregated network current and network voltage.

[0011] On the other hand, before determining the weights of each training sample point for training the proxy models to be aggregated, it further includes:

[0012] Obtain multiple proxy models;

[0013] Perform validation error on each of the proxy models to obtain an error value;

[0014] Select a preset number of proxy models with the lowest error values as the proxy models to be aggregated.

[0015] On the other hand, performing validation error on each of the proxy models to obtain an error value includes:

[0016] Each time, omit one training sample point, construct the proxy model, and obtain the predicted value output by the proxy model;

[0017] Determine the compensated cross-validation error of each proxy model according to the compensated cross-validation error relationship formula;

[0018] The compensated cross-validation error relationship formula is:

[0019] ;

[0020] where CCV m is the compensated cross-validation error of the mth proxy model, n is the total number of training sample points, M is the total number of proxy models, y k is the true response value of the kth training sample point, is the predicted value of the mth proxy model at the kth training sample point without the jth training sample point, y j is the true response value of the jth training sample point, is the average of the true responses of the training samples.

[0021] On the other hand, determining the weights of each training sample point for training the proxy models to be aggregated includes:

[0022] Taking the minimum difference between the predicted value of the training sample point output by each proxy model to be aggregated based on the input training sample point and the true response value of the training sample point as the goal, construct an objective function;

[0023] The expression of the objective function is:

[0024] ;

[0025] The constraint conditions are:

[0026] ;

[0027] Determine the weights of each training sample point based on the objective function;

[0028] where, w mi is the weight of the m-th surrogate model at the i-th training sample point x i , is the predicted value of the i-th training sample point x i output by the m-th surrogate model when omitting the i-th training sample point x i , y i is the true response value of the i-th training sample point.

[0029] On the other hand, after determining the weights of each training sample point for training the surrogate models to be aggregated, it further includes:

[0030] Judge whether the prediction point belongs to the interpolation region;

[0031] If , it is determined that the prediction point belongs to the interpolation region, and enter the step of determining the weights of the prediction points in the interpolation region under each surrogate model to be aggregated according to the weights of each training sample point;

[0032] If , it is determined that the prediction point belongs to the extrapolation region, and enter the step of determining the weights of the prediction points in the extrapolation region under each surrogate model to be aggregated according to the accuracy of each surrogate model to be aggregated;

[0033] where, is the p-th prediction point, is the minimum value of the training sample points, is the maximum value of the training sample points.

[0034] On the other hand, determining the weights of the prediction points in the interpolation region under each surrogate model to be aggregated according to the weights of each training sample point includes:

[0035] Determine the weights of the prediction points in the interpolation region under each surrogate model to be aggregated according to the weights of each training sample point, and the expression of the weights of the prediction points in the interpolation region under each surrogate model to be aggregated is:

[0036] ;

[0037] Among them, is the weight at the i-th prediction point within the interpolation region under the m-th surrogate model, n is the total number of training sample points, and w mj is the weight of the m-th surrogate model at the j-th training sample point x j , d ij is the i-th prediction point and the Euclidean norm between the j-th training point x j , w mk is the weight of the m-th surrogate model at the k-th training sample point x k , d jk is the Euclidean norm between the j-th training sample point x j and the k-th training sample point x k .

[0038] On the other hand, after determining the weights of the prediction points within the interpolation regions under the respective surrogate models to be aggregated according to the weights of the respective training sample points, it further includes:

[0039] Normalize the weights of the prediction points within the interpolation region. The expression for the weights of the prediction points within the normalized interpolation region is:

[0040] ;

[0041] Among them, is the weight at the i-th prediction point within the interpolation region under the m-th surrogate model after normalization, .

[0042] On the other hand, determining the weights of the prediction points within the extrapolation regions under the respective surrogate models to be aggregated according to the accuracies of the respective surrogate models to be aggregated includes:

[0043] Determining the weights of the prediction points within the extrapolation regions under the respective surrogate models to be aggregated according to the accuracies of the respective surrogate models to be aggregated. The expression for the weights of the prediction points within the extrapolation regions under the respective surrogate models to be aggregated is:

[0044] ;

[0045] Among them, is the weight of the m-th surrogate model to be aggregated in the extrapolation region, E m is the compensated cross-validation error of the m-th surrogate model to be aggregated, and the compensated cross-validation error is negatively correlated with the accuracy, is the average value of the compensated cross-validation errors of all surrogate models to be aggregated, and both α and β are hyperparameters, .

[0046] On the other hand, aggregating the responses output by multiple agent models to be aggregated based on the weights of the prediction points in the interpolation region under each agent model to be aggregated, the responses of the prediction points in the interpolation region output by each agent model to be aggregated, the weights of the prediction points in the extrapolation region under each agent model to be aggregated, and the responses of the prediction points in the extrapolation region output by each agent model to be aggregated includes:

[0047] Aggregating the responses output by multiple agent models to be aggregated based on the weights of the prediction points in the interpolation region under each agent model to be aggregated, the responses of the prediction points in the interpolation region output by each agent model to be aggregated, the weights of the prediction points in the extrapolation region under each agent model to be aggregated, and the responses of the prediction points in the extrapolation region output by each agent model to be aggregated, and the relational expression for the aggregation is:

[0048] ;

[0049] Wherein, is the response of the prediction point x output after aggregating each agent model to be aggregated, is the response of the prediction point x output by the m-th agent model to be aggregated. When the prediction point x belongs to the interpolation region, w m (x) is , is the weight of the i-th prediction point in the interpolation region under the m-th agent model. When the prediction point x belongs to the extrapolation region, w m (x) is , is the weight of the m-th agent model to be aggregated in the extrapolation region.

[0050] To solve the above technical problems, the present invention also provides a pantograph control device, including:

[0051] A memory for storing a computer program;

[0052] A processor for implementing the steps of the above pantograph control method when executing the computer program.

[0053] The present application provides a method and device for controlling a pantograph, which relates to the field of data fusion, and includes: determining the weights of the prediction points in the interpolation region under each proxy model to be aggregated according to the weights of each training sample point; determining the weights of the prediction points in the extrapolation region under each proxy model to be aggregated according to the accuracy of each proxy model to be aggregated; aggregating the responses output by multiple proxy models to be aggregated based on the weights of the prediction points in the interpolation region and the output responses and the weights of the prediction points in the extrapolation region and the output responses. Assuming that the weight of the prediction point is proportional to the gradient of the weight of the training sample point, the weight of the prediction point in the interpolation region is determined based on the weight of the training sample point. Since the extrapolation region is not trained, the weight is determined based on the accuracy of the proxy model. By assigning weights to different proxy models, the evaluation between multiple models and the adjustment of the weight ratio are better carried out, and the network current and network voltage of the pantograph output are more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the prior art and the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0055] Figure 1 is a flowchart of a method for controlling a pantograph provided by the present invention;

[0056] Figure 2 is a flowchart of another method for controlling a pantograph provided by the present invention;

[0057] Figure 3 is a schematic structural diagram of a device for controlling a pantograph provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] The core of the present invention is to provide a method and device for controlling a pantograph. By assigning weights to different proxy models, the evaluation between multiple models and the adjustment of the weight ratio are better carried out, and the network current and network voltage of the pantograph output are more accurate.

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0060] Figure 1The flowchart of a control method for a pantograph provided by the present invention. The control method for the pantograph includes:

[0061] S11: Determine each training sample point and each prediction point. Each prediction point and each training sample point include at least two of the pantograph lifting force, train running speed, environmental temperature, and catenary suspension height.

[0062] S12: Determine the weights of each training sample point for training the proxy models to be aggregated.

[0063] S13: Determine the weights of the prediction points within the interpolation region under each proxy model to be aggregated according to the weights of each training sample point. The range where the training sample points are located is the interpolation region.

[0064] S14: Determine the weights of the prediction points within the extrapolation region under each proxy model to be aggregated according to the accuracy of each proxy model to be aggregated. The range outside the range where the training sample points are located is the extrapolation region.

[0065] S15: Aggregate the responses output by multiple proxy models to be aggregated based on the weights of the prediction points within the interpolation region under each proxy model to be aggregated, the responses of the prediction points within the interpolation region output by each proxy model to be aggregated, the weights of the prediction points within the extrapolation region under each proxy model to be aggregated, and the responses of the prediction points within the extrapolation region output by each proxy model to be aggregated. The response is the catenary current and catenary voltage of the pantograph.

[0066] S16: Output the aggregated catenary current and catenary voltage.

[0067] In high-speed railways or urban rail transit, the pantograph is the core device for trains to obtain electrical energy from the catenary. Its dynamic performance directly affects the stability and safety of power transmission. Among them, the catenary current and catenary voltage are key parameters. However, the interaction between the pantograph and the catenary is affected by multiple factors (such as lifting force, running speed, environmental temperature, catenary suspension height, etc.). Traditional single models are difficult to accurately predict the catenary current and catenary voltage under complex working conditions. Through the multi-agent model integration technology (POEM), the advantages of different models can be dynamically fused to improve the prediction accuracy and robustness.

[0068] The training sample points include: pantograph lifting force (kN), train running speed (km / h), environmental temperature (°C), and catenary suspension height (mm). The prediction points are the pantograph lifting force (kN), train running speed (km / h), environmental temperature (°C), and catenary suspension height (mm) under new working conditions. The true response is the measured catenary current (A) and measured catenary voltage (kV) corresponding to the training sample points. What is finally output is the catenary current and catenary voltage predicted based on the prediction points.

[0069] Table 1 is the training data table input to each surrogate model, including training sample points and true responses.

[0070] Table 2 is the data table of prediction points input to each surrogate model.

[0071] Table 3 is the network flow and network voltage table output after aggregation of each surrogate model to be aggregated.

[0072] Table 1

[0073]

[0074] Table 2

[0075]

[0076] Table 3

[0077]

[0078] Since each surrogate model has its own advantages and disadvantages for different problems, no single surrogate model always performs best on all problems. To simultaneously obtain the advantages of multiple models, multiple surrogate models can be combined in an integrated form to obtain excellent and more reliable prediction performance. Compared with a single model, the method of integrating multiple surrogate models has higher prediction accuracy and robustness. Usually, the method of determining the weighting coefficient is average integration, that is, the method of using global metrics to determine the weight coefficient of each surrogate model. In such a combination, the weight coefficient remains fixed throughout the design space. However, using a constant weight coefficient combination cannot fully utilize the local characteristics of each surrogate model, thus limiting the further improvement of prediction accuracy.

[0079] Therefore, this application considers the weights of prediction points from two perspectives, namely whether the prediction point belongs to the interpolation region. It can be understood that the range where the training sample points are located is the interpolation region. The input data within the interpolation region has been trained, and the results output by the surrogate model are also relatively accurate. However, the extrapolation region is data that has not been trained by the training sample points, and the results output by the surrogate model are inaccurate for the surrogate model. Therefore, in the interpolation region, this application adopts the method that the weighting coefficient at the prediction point is proportional to the gradient of the weighting coefficient at the training point, that is, the weight of the training sample point can be used to derive the weight of the prediction point. In the extrapolation region, it is determined based on the accuracy of the surrogate model to be aggregated. Then, different weights are divided based on the region where the prediction point is located, and the finally obtained network flow and network voltage are more accurate.

[0080] Further, the multi-agent model fusion method provided in this application can also be applied to the vibration signal prediction of vehicle bogies or the prediction of vehicle running noise. This solution can be used for data affected by various factors. This application only takes the network current and network voltage of vehicle pantographs as examples.

[0081] This application provides a control method for pantographs, which relates to the field of data fusion and includes: determining the weights of prediction points within the interpolation region under each agent model to be aggregated according to the weights of each training sample point; determining the weights of prediction points within the extrapolation region under each agent model to be aggregated according to the accuracy of each agent model to be aggregated; aggregating the responses output by multiple agent models to be aggregated based on the weights of the prediction points within the interpolation region and the output responses and the weights of the prediction points within the extrapolation region and the output responses. Assuming that the weight of the prediction point is proportional to the gradient of the weight of the training sample point, the weight of the prediction point within the interpolation region is determined based on the weight of the training sample point. Since the extrapolation region has not been trained, the weight is determined based on the accuracy of the agent model. By assigning weights to different agent models, better evaluation and weight ratio adjustment between multiple models can be carried out, and the network current and network voltage of the pantograph output are more accurate.

[0082] Based on the above embodiments:

[0083] In some embodiments, before determining the weights of each training sample point for training the agent models to be aggregated, it further includes:

[0084] Obtaining multiple agent models;

[0085] Performing verification errors on each agent model to obtain error values;

[0086] Selecting a preset number of agent models with the lowest error values as the agent models to be aggregated.

[0087] Considering that different agent models have different effects on non-linear problems in different dimensions, multiple agent models are initially selected as the candidate model set. Before constructing the integrated model, it is necessary to evaluate the accuracy of individual models and filter out individual models with poor accuracy. Here, we use the CCV (compensated cross-validation error) error to evaluate the performance of individual agent models.

[0088] In some embodiments, performing verification errors on each agent model to obtain error values includes:

[0089] Each time, omit one training sample point, construct the agent model, and obtain the predicted value output by the agent model;

[0090] Determining the compensated cross-validation error of each agent model according to the compensated cross-validation error relationship formula;

[0091] The compensation cross - validation error relationship is as follows:

[0092] ;

[0093] where CCV m is the compensation cross - validation error of the m - th surrogate model, n is the total number of training sample points, M is the total number of surrogate models, y k is the true response value of the k - th training sample point, is the predicted value of the k - th training sample point by the m - th surrogate model without the j - th training sample point, y j is the true response value of the j - th training sample point, is the average value of the true responses of the training samples.

[0094] In the proposed CCV, it is necessary to construct multiple surrogate models multiple times, each time omitting one training sample point. Finally, select multiple surrogate models with the smallest compensation cross - validation error as the surrogate models to be aggregated.

[0095] In some embodiments, determining the weights of the respective training sample points for training the surrogate models to be aggregated includes:

[0096] Taking the minimum difference between the predicted value of the training sample point output by each surrogate model to be aggregated based on the input training sample points and the true response value of the training sample point as the objective, construct an objective function;

[0097] The expression of the objective function is:

[0098] ;

[0099] The constraint conditions are:

[0100] ;

[0101] Determine the weights of the respective training sample points based on the objective function;

[0102] where w mi is the weight of the m - th surrogate model at the i - th training sample point x i , is the predicted value of the i - th training sample point x i output by the m - th surrogate model when omitting the i - th training sample point x i , y i is the true response value of the i - th training sample point.

[0103] The key to constructing the multi-model aggregation lies in selecting a suitable set of point-wise weight coefficients for each surrogate model. Given the true responses of the omitted training sample points, calculating the weight coefficient for a point can be expressed as an optimization problem. The aim of this optimization problem is to find a set of weight coefficients such that the predicted point equals the true response at that point. An optimization problem is constructed for each training sample point:

[0104] ;

[0105] where w mi is greater than or equal to 0, and the sum of the weights of each surrogate model is 1.

[0106] Assume that the predictions of all individual surrogate models at the omitted training points are greater than or less than the true responses. In this case, it is easy to infer that the weight factor of the surrogate model with the prediction closest to the true response should be 1, while the weight factors of the remaining surrogate models are 0. Otherwise, the weight factors are calculated according to the optimization problem. Figure 2 In the flowchart of ", the " " represents the iterative process when calculating the weights of training sample points. Specifically, each training sample point (e.g., the j-th point) is successively omitted, the surrogate model is retrained based on the remaining samples, and the prediction error (CCV) of the omitted point is calculated. This process is looped (i is the iteration counter) until all training sample points have been traversed once. The goal of the optimization problem is to find a set of weight coefficients for each omitted sample point such that the predicted value of the ensemble model is as close as possible to the true response. The iterative process ( ) corresponds to the weight optimization for each sample point, and finally, the contribution weights of each surrogate model in different local regions are screened out. Figure 2 The number of factors in

[0107] In some embodiments, after determining the weights of each training sample point for the surrogate models to be aggregated, it further includes:

[0108] Determine whether the prediction point belongs to the interpolation region;

[0109] If , it is determined that the prediction point belongs to the interpolation region, and proceed to the step of determining the weights of the prediction points in the interpolation region for each surrogate model to be aggregated according to the weights of each training sample point;

[0110] If , it is determined that the prediction point belongs to the extrapolation region, and proceed to the step of determining the weights of the prediction points in the extrapolation region for each surrogate model to be aggregated according to the accuracy of each surrogate model to be aggregated;

[0111] where is the p-th prediction point, is the minimum value of the training sample points, is the maximum value of the training sample points.

[0112] Since the experimental design does not always sample at the boundaries of the design space, this method defines the area within the sampling range as the interpolation area and the remaining area as the extrapolation area. If it is in the interpolation area, it is considered that the response corresponding to the prediction point can be more accurately output by the proxy model to be aggregated. If it is in the extrapolation area, it is considered that the response corresponding to the prediction point output by the proxy model to be aggregated is not accurate enough.

[0113] In some embodiments, after determining the weights of the prediction points in the interpolation area under each proxy model to be aggregated according to the weights of each training sample point, it further includes:

[0114] Determining the weights of the prediction points in the interpolation area under each proxy model to be aggregated according to the weights of each training sample point, and the expression of the weights of the prediction points in the interpolation area under each proxy model to be aggregated is:

[0115] ;

[0116] where, is the weight of the i-th prediction point in the interpolation area under the m-th proxy model, n is the total number of training sample points, is the weight of the m-th proxy model at the j-th training sample point x j , d ij is the Euclidean norm between the i-th prediction point and the j-th training point x j , is the weight of the m-th proxy model at the k-th training sample point x k , d jk is the Euclidean norm between the j-th training sample point x j and the k-th training sample point x k .

[0117] In this paper, the ratio between the prediction point and the training point is obtained through the distance between the prediction point and the training point and the weight coefficient of the training point. Since the gradient of the weight coefficient at the training point can reflect the change trend of the weight in the entire design space, in the interpolation area, we assume that the weighted coefficient at the prediction point is proportional to the gradient of the weight coefficient at the training point:

[0118] .

[0119] In some embodiments, after determining the weights of the prediction points in the interpolation area under each proxy model to be aggregated according to the weights of each training sample point, it further includes:

[0120] Normalize the weights of the prediction points in the interpolation region. The expression for the weights of the prediction points in the interpolation region after normalization is:

[0121] ;

[0122] where, is the weight of the i-th prediction point in the interpolation region under the m-th surrogate model after normalization, .

[0123] Since the gradient of the weight coefficient may be negative, the calculated weights cannot be directly used to calculate the response of the ensemble model. Here, we make corresponding adjustments to it. First, normalize the weights, and then adjust the negative weights to 0.

[0124] In some embodiments, determining the weights of the prediction points in the extrapolation region under each surrogate model to be aggregated according to the accuracy of each surrogate model to be aggregated includes:

[0125] Determine the weights of the prediction points in the extrapolation region under each surrogate model to be aggregated according to the accuracy of each surrogate model to be aggregated. The expression for the weights of the prediction points in the extrapolation region under each surrogate model to be aggregated is:

[0126] ;

[0127] where, is the weight of the m-th surrogate model to be aggregated in the extrapolation region, E m is the compensated cross-validation error of the m-th surrogate model to be aggregated. The compensated cross-validation error is negatively correlated with the accuracy, is the average of the compensated cross-validation errors of all surrogate models to be aggregated. Both α and β are hyperparameters, , .

[0128] Since the extrapolation region is data that has not been trained, it needs to be determined based on the accuracy of the data output by the surrogate models to be aggregated themselves. Therefore, this application uses the compensated cross-validation error to characterize the accuracy of the surrogate models to be aggregated, and then determines the weights of the surrogate models to be aggregated in the extrapolation region based on this.

[0129] In this application, take , .

[0130] In some embodiments, aggregating multiple surrogate models to be aggregated based on the weights of the prediction points in the interpolation region under each surrogate model to be aggregated, the responses of the prediction points in the interpolation region output by each surrogate model to be aggregated, the weights of the prediction points in the extrapolation region under each surrogate model to be aggregated, and the responses of the prediction points in the extrapolation region output by each surrogate model to be aggregated includes:

[0131] Aggregate multiple agent models to be aggregated based on the weights of the prediction points in the interpolation region under each agent model to be aggregated, the responses of the prediction points in the interpolation region output by each agent model to be aggregated, the weights of the prediction points in the extrapolation region under each agent model to be aggregated, and the responses of the prediction points in the extrapolation region output by each agent model to be aggregated. The aggregation relationship is as follows:

[0132] ;

[0133] Among them, is the response of the prediction point x output after aggregating each agent model to be aggregated, is the response of the prediction point x output by the m-th agent model to be aggregated. When the prediction point x belongs to the interpolation region, w m (x) is , is the weight of the i-th prediction point in the interpolation region under the m-th agent model. When the prediction point x belongs to the extrapolation region, w m (x) is , is the weight of the m-th agent model to be aggregated in the extrapolation region. For any input point, the predicted value of the integrated model is calculated according to . w m (x) is dynamically selected according to the region where the prediction point is located. The interpolation region uses , and the extrapolation region uses .

[0134] Figure 3 FIG.

[0135] Memory 31, which is used to store computer programs;

[0136] Processor 32, which is used to implement the steps of the above-mentioned pantograph control method when executing the computer program.

[0137] For the introduction of the pantograph control device provided in this application, please refer to the above-mentioned embodiments, and details will not be repeated here.

[0138] It should also be noted that in this specification, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variation thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0139] Those skilled in the art can further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present invention.

[0140] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A control method for a pantograph, characterized in that, Including: Determine each training sample point and each prediction point. Each of the prediction points and each of the training sample points includes at least two of the pantograph lift force, train running speed, ambient temperature, and catenary suspension height; Determine the weights of each training sample point for training the proxy models to be aggregated; Determine the weights of the prediction points within the interpolation region under each proxy model to be aggregated according to the weights of each training sample point. The range where the training sample points are located is the interpolation region; Determine the weights of the prediction points within the extrapolation region under each proxy model to be aggregated according to the accuracy of each of the proxy models to be aggregated. The range outside the range where the training sample points are located is the extrapolation region; Aggregate the responses output by multiple proxy models to be aggregated based on the weights of the prediction points within the interpolation region under each of the proxy models to be aggregated, the responses of the prediction points within the interpolation region output by each of the proxy models to be aggregated, the weights of the prediction points within the extrapolation region under each of the proxy models to be aggregated, and the responses of the prediction points within the extrapolation region output by each of the proxy models to be aggregated. The response is the net current and net voltage of the pantograph; Output the aggregated net current and net voltage.

2. The control method of the pantograph according to claim 1, characterized in that, Before determining the weights of each training sample point for training the proxy models to be aggregated, it further includes: Obtain multiple proxy models; Perform verification errors on each of the proxy models to obtain error values; Select a preset number of proxy models with the lowest error values as the proxy models to be aggregated.

3. The control method of the pantograph according to claim 2, characterized in that Performing verification errors on each of the proxy models to obtain error values includes: Each time, omit one training sample point, construct the proxy model, and obtain the predicted value output by the proxy model; Determine the compensated cross-validation error of each proxy model according to the compensated cross-validation error relation formula; The compensated cross-validation error relation formula is: ； Among them, CCV m is the compensated cross - validation error of the m - th surrogate model, n is the total number of training sample points, M is the total number of surrogate models, y k is the true response value of the k - th training sample point, is the predicted value of the m - th surrogate model at the k - th training sample point without the j - th training sample point, y j is the true response value of the j - th training sample point, is the average value of the true responses of the training samples.

4. The control method of the pantograph according to claim 1, characterized in that Determining the weights of each training sample point for training the proxy models to be aggregated includes: Construct an objective function with the goal of minimizing the difference between the predicted value of the training sample point output by each proxy model to be aggregated based on the input training sample point and the true response value of the training sample point; The expression of the objective function is: ； The constraint condition is: ； Determine the weights of each training sample point based on the objective function; where, w mi is the weight of the m-th surrogate model at the i-th training sample point x i , is the predicted value of the i-th training sample point x i output by the m-th surrogate model when omitting the i-th training sample point x i , and y i is the true response value of the i-th training sample point.

5. The control method of the pantograph according to claim 1, characterized in that, After determining the weights of each training sample point for training the proxy models to be aggregated, it further includes: Judge whether the prediction point belongs to the interpolation region; If , it is determined that the prediction point belongs to the interpolation region, and the process proceeds to the step of determining the weights of the prediction points within the interpolation region under each aggregation proxy model according to the weights of the respective training sample points; If , it is determined that the predicted point belongs to the extrapolation region, and the step of determining the weights of the predicted points in the extrapolation region under each agent model to be aggregated according to the accuracy of each agent model to be aggregated is entered; Among them, is the p-th prediction point, is the minimum value of the training sample points, is the maximum value of the training sample points.

6. The control method of the pantograph according to claim 1, characterized in that Determining the weights of the prediction points within the interpolation region under each proxy model to be aggregated according to the weights of each training sample point includes: Determine the weights of the prediction points within the interpolation region under each proxy model to be aggregated according to the weights of each training sample point. The expression of the weights of the prediction points within the interpolation region under each proxy model to be aggregated is: ； Among them, is the weight at the i-th prediction point in the interpolation region under the m-th surrogate model, n is the total number of training sample points, and w mj is the weight of the m-th surrogate model at the j-th training sample point x j , d ij is the i-th prediction point and the Euclidean norm between the j-th training point x j . w mk is the weight of the m-th surrogate model at the k-th training sample point x k , d jk is the Euclidean norm between the j-th training sample point x j and the k-th training sample point x k .

7. The control method of the pantograph according to claim 6, characterized in that, After determining the weights of the prediction points within the interpolation region under each proxy model to be aggregated according to the weights of each training sample point, it further includes: Normalize the weights of the prediction points within the interpolation region. The expression of the normalized weights of the prediction points within the interpolation region is: ； Among them, is the weight at the i-th prediction point in the interpolation region under the m-th surrogate model after normalization, .

8. The control method of the pantograph according to claim 1, characterized in that, Determining the weights of the prediction points within the extrapolation region under each proxy model to be aggregated according to the accuracy of each of the proxy models to be aggregated includes: Determine the weights of the prediction points in the extrapolation region under each of the to-be-aggregated surrogate models according to the accuracy of each of the to-be-aggregated surrogate models, and the expression of the weights of the prediction points in the extrapolation region under each of the to-be-aggregated surrogate models is: ； Among them, is the weight of the m-th proxy model to be aggregated in the extrapolation region, E m is the compensated cross-validation error of the m-th proxy model to be aggregated, and the compensated cross-validation error is negatively correlated with accuracy. is the average value of the compensated cross-validation errors of all proxy models to be aggregated. Both α and β are hyperparameters, where α < 1 and β < 0.

9. The control method of the pantograph according to claim 1, characterized in that, Aggregate the responses output by multiple to-be-aggregated surrogate models based on the weights of the prediction points in the interpolation region under each of the to-be-aggregated surrogate models, the responses of the prediction points in the interpolation region output by each of the to-be-aggregated surrogate models, the weights of the prediction points in the extrapolation region under each of the to-be-aggregated surrogate models, and the responses of the prediction points in the extrapolation region output by each of the to-be-aggregated surrogate models, including: Aggregate the responses output by multiple to-be-aggregated surrogate models based on the weights of the prediction points in the interpolation region under each of the to-be-aggregated surrogate models, the responses of the prediction points in the interpolation region output by each of the to-be-aggregated surrogate models, the weights of the prediction points in the extrapolation region under each of the to-be-aggregated surrogate models, and the responses of the prediction points in the extrapolation region output by each of the to-be-aggregated surrogate models, and the relationship formula of the aggregation is: ； Among them, is the response of the predicted point x output after aggregating each agent model to be aggregated, is the response of the predicted point x output by the m-th agent model to be aggregated. When the predicted point x belongs to the interpolation region, w m (x) is , is the weight of the i-th predicted point in the interpolation region under the m-th agent model. When the predicted point x belongs to the extrapolation region, w m (x) is , is the weight of the m-th agent model to be aggregated in the extrapolation region.

10. A control device for a pantograph, characterized in that, Including: A memory for storing a computer program; A processor for implementing the steps of the pantograph control method according to any one of claims 1 to 9 when executing the computer program.

Images