New energy automobile remaining mileage prediction method based on neural network

Abstract

A new energy automobile remaining mileage prediction method based on a neural network relates to the field of new energy automobiles, designs a bidirectional recurrent neural network BRNN, and can extract features from multiple time dimensions and capture dynamic change characteristics of time series data, so that a prediction model can more accurately reflect energy consumption changes in an actual driving process, and the prediction accuracy is improved. And the residual mileage prediction precision is improved. A multi-reservoir echo state network DAMESN model based on a zoom dot product attention mechanism is adopted to predict the remaining driving mileage, and a complex nonlinear relation is efficiently captured in combination with a depth feature vector and the zoom dot product attention mechanism, so that the generalization ability and the prediction precision of the model are enhanced, and meanwhile, the driving experience is optimized; a driver is helped to more accurately plan the journey, and the risk of electric quantity exhaustion is reduced.

Description

Technical Field

[0001] The present invention relates to the field of new energy vehicles, and in particular to a method for predicting the remaining mileage of a new energy vehicle based on a neural network. Background Art

[0002] With the popularity of new energy vehicles, users need to accurately obtain the remaining mileage prediction of new energy vehicles through the car computer to better plan the charging time and charging location. However, the remaining mileage prediction of new energy vehicles in the existing technology has the following disadvantages:

[0003] 1) Dynamic time series data is not fully considered; 2) External influencing factors (temperature) cannot be included in the prediction; 3) The intelligence level of electric vehicle remaining mileage prediction is low.

[0004] The Chinese invention patent with application number CN202210127495.9 proposes a remaining mileage prediction method, device, electric vehicle and storage medium, which relates to the field of automobile technology. It solves the problem of accurate and efficient prediction of the remaining driving mileage of electric vehicles under specific road conditions, specific vehicles and specific driving habits. The method includes: at each sampling moment, obtaining the motor torque, motor rotation angular velocity, first speed and battery charge state of the electric vehicle at the current moment; determining the operating mode of the electric vehicle at the current moment according to the braking torque value and the motor torque of the electric vehicle at the current moment, and determining the average power of the motor torque, average speed and the remaining battery power value at the previous k moments according to the operating mode; determining the remaining driving time and remaining driving mileage according to the remaining battery power at the previous k moments, the average power of the motor torque at the previous k moments, and the average speed at the previous k moments, where k is a positive integer.

[0005] The disadvantages of this technology are: 1. This technology directly predicts the remaining mileage based on the average data of historical moments, which ignores the dynamic change characteristics of time series data, resulting in the inability to fully extract and utilize the features of time series data; 2. This technology only considers the influence of the vehicle's own parameters, without considering the influence of the external environment, and to a certain extent cannot fully guarantee the accuracy of power prediction.

[0006] The Chinese invention patent with application number CN202210716648.3 proposes a method, device, electronic device and vehicle for predicting the remaining mileage of a vehicle, which relates to the field of data processing technology, wherein the method includes: obtaining the remaining battery power of the vehicle; if the remaining battery power is less than or equal to a preset power threshold, obtaining the energy consumption per kilometer of the vehicle, and obtaining the remaining battery energy of the vehicle; dividing the remaining battery energy by the energy consumption per kilometer to obtain the predicted value of the pure electric remaining mileage of the vehicle. By applying the technical solution of this application, when the extended-range electric vehicle is in the low SOC segment, the real-time pure electric remaining mileage value can be accurately predicted, which improves the prediction accuracy of the pure electric remaining mileage in the low SOC segment, and the more accurate pure electric remaining mileage prediction value can better help the driver control the car and improve the car's control experience.

[0007] The disadvantages of this technology are: 1. This technology directly calculates the remaining mileage prediction value of the electric vehicle by dividing the remaining battery power of the vehicle by the energy consumption per kilometer. However, the remaining mileage of the electric vehicle is affected by complex factors, and the prediction result obtained by direct calculation is less accurate; 2. This technology uses the remaining battery energy and a fixed energy consumption ratio to predict the mileage, ignoring the changes in battery performance over time and usage. For example, the performance of the battery will be different at different temperatures, and as the battery ages, its capacity and discharge efficiency will also change. Ignoring these factors may lead to deviations in the prediction results. Summary of the invention

[0008] In order to overcome the deficiencies of the above technologies, the present invention provides a method for predicting the remaining mileage by fully considering the influence of various complex factors on the remaining mileage and making full use of the dynamic and interactive characteristics of mileage time series data to achieve a more accurate remaining mileage prediction.

[0009] The technical solution adopted by the present invention to overcome the technical problems is:

[0010] A method for predicting the remaining mileage of a new energy vehicle based on a neural network, characterized in that it comprises the following steps: a) collecting N groups of input feature data of electric vehicles to obtain an input feature data set X, where X = {X1, X2, ..., X i ,...,X N}, X i For the i-th group of input feature data, i∈{1,...,N}, the remaining mileage corresponding to the i-th group of input feature data is Y i ;

[0011] b) Establish a feature extraction module and input the feature data X of the i-th group i Input into the feature extraction module, and output the final extracted deep feature vector F i ;

[0012] c) Establish a multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism, and transform the deep feature vector F i The input is fed into the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism, and the remaining mileage prediction value is output.

[0013] d) Using the remaining mileage prediction value And the remaining mileage Y i Train the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism to obtain the optimized multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism;

[0014] e) The optimized multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is deployed into the vehicle control module of the electric vehicle.

[0015] Further, step a) comprises the following steps:

[0016] a-1) Collecting data of electric vehicles once a minute, including: vehicle speed, vehicle load current, ambient temperature, remaining battery capacity, and current mileage of the vehicle;

[0017] a-2) The characteristic data of the electric vehicle collected every 30 times constitutes a set of input characteristic data, X i =[V i ,I i ,T i ,C i ],V i is the speed of the car with the i-th set of input feature data, V i =[v1,v2,...,v i ,...,v 30 ],v i is the speed of the car collected at the i-th minute, I i The vehicle load current of the i-th group of input characteristic data, I i =[e1,e2,...,e i ,...,e 30 ],e i is the vehicle load current collected at the i-th minute, T i is the ambient temperature of the i-th group of input characteristic data, T i =[t1,t2,...,t i ,...,t 30 ], t i is the ambient temperature collected at the i-th minute, C i is the remaining battery capacity of the i-th set of input feature data, Ci =[c1,c2,...,c i ,...,c 30 ], c i is the remaining battery capacity collected at the i-th minute;

[0018] a-3) Record the final mileage of the vehicle when the battery is exhausted, and the current mileage of the vehicle collected at the i-th minute a i , subtract the current cruising range of the vehicle collected at the i-th minute from the final cruising range mileage a i Get the remaining mileage y in the i-th minute i , Y i =[y1,y2,...,y i ,...,y 30 ].

[0019] Further, step b) comprises the following steps:

[0020] b-1) The feature extraction module is a bidirectional recurrent neural network BRNN;

[0021] b-2) Input the feature data X of the i-th group i The data in the image are input into the two forward RNN layers and convolutional layers of the bidirectional recurrent neural network BRNN in the forward order from time 1 to 30, and the input is the feature X i ′, the i-th group input feature data X i The data in the image are input into the two reverse RNN layers and convolutional layers of the bidirectional recurrent neural network BRNN in reverse order from time 30 to 1, and the output is the feature X. i ″, feature X i ' and characteristic X i ″ is sequentially input into the fully connected layer and Dropout layer of the bidirectional recurrent neural network BRNN, and the final extracted deep feature vector F is output i .

[0022] Further, step c) comprises the following steps:

[0023] c-1) The multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism consists of an input layer, an internal reservoir unit, a dot product scaled attention mechanism, and an output layer;

[0024] c-2) Deep feature vector F i The dimension is Q. The input layer of the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism consists of Q neurons. i Each dimension of is input into the corresponding neuron of the input layer, and the output is the feature vector Fi ′;

[0025] c-3) The internal reservoir unit of the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism consists of M internal storage layers, each of which consists of P sparsely connected neurons. i ′ is input into M internal storage layers in sequence, and the output is the feature vector F i ″;

[0026] c-4) The dot product scaling attention mechanism of the multi-reservoir echo state network model DAMESN based on the scaling dot product attention mechanism is composed of a linear layer, a first dot product layer, a scaling layer, a softmax activation function, and a second dot product layer. i ″ is sequentially input into the linear layer of the dot product scaling attention mechanism, the first dot product layer for dot product calculation, the scaling layer, the softmax activation function, and the second dot product layer for dot product calculation, and the feature vector F is output. i ″′;

[0027] c-5) The output layer of the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is composed of a softmax activation function, which converts the feature vector F i ″′ is input to the output layer, and the remaining mileage prediction value is output

[0028] Further, step d) comprises the following steps:

[0029] d-1) Through the formula Calculate the objective function f;

[0030] d-2) Using the Bayesian optimization method, the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is trained based on the objective function f to obtain the optimized multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism.

[0031] The beneficial effects of the present invention are as follows: in the process of data collection and feature extraction, four factors, namely ambient temperature, vehicle speed, load current and remaining battery capacity, are incorporated, and the model can reflect changes in energy consumption under complex driving conditions, further improving the reliability and accuracy of the prediction, and making the model more adaptable under different driving conditions. A bidirectional recurrent neural network BRNN is designed, which can extract features from multiple time dimensions and capture the dynamic change characteristics of time series data, so that the prediction model can more accurately reflect changes in energy consumption during actual driving, and improve the accuracy of remaining mileage prediction. The multi-reservoir echo state network DAMESN model based on the scaled dot product attention mechanism is used to predict the remaining mileage, combining deep feature vectors and scaled dot product attention mechanisms to efficiently capture complex nonlinear relationships, enhancing the generalization ability and prediction accuracy of the model while optimizing the driving experience, helping drivers to plan their trips more accurately and reduce the risk of running out of power. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flow chart of the method of the present invention;

[0033] Figure 2 A network structure diagram of a feature extraction module of the present invention;

[0034] Figure 3 This is a structural diagram of a multi-reservoir echo state network model based on a scaled dot product attention mechanism of the present invention;

[0035] Figure 4 It is a structural diagram of the remaining mileage prediction model of the present invention. DETAILED DESCRIPTION

[0036] The following is combined with Figure 1 To Attachment Figure 4 The present invention is further described.

[0037] A method for predicting the remaining mileage of a new energy vehicle based on a neural network, characterized in that it comprises the following steps: a) collecting N groups of input feature data of electric vehicles to obtain an input feature data set X, where X = {X1, X2, ..., X i ,...,X N}, X i For the i-th group of input feature data, i∈{1,...,N}, the remaining mileage corresponding to the i-th group of input feature data is Y i .

[0038] b) Establish a feature extraction module and input the feature data X of the i-th group i Input into the feature extraction module, and output the final extracted deep feature vector F i .

[0039] c) Establish a multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism, and transform the deep feature vector F i The input is fed into the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism, and the remaining mileage prediction value is output.

[0040] d) Using the remaining mileage prediction value And the remaining mileage Y i The multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is trained to obtain the optimized multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism.

[0041] e) The optimized multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is deployed in the vehicle control module of the electric vehicle to detect the vehicle speed, ambient temperature, vehicle load current and battery remaining capacity information in real time as input feature data. Then, based on the feature extraction module and the remaining mileage prediction model, the prediction result of the electric vehicle's cruising range is obtained in real time.

[0042] The following problems of the prior art are solved: 1) dynamic time series data is not fully considered; 2) external influencing factors (temperature) cannot be included in the prediction; 3) the intelligence level of the remaining mileage prediction of electric vehicles is low. The present invention fully considers the influence of various complex factors on the remaining mileage, and makes full use of the dynamic and interactive characteristics of mileage time series data to achieve a more accurate remaining mileage prediction.

[0043] In one embodiment of the present invention, step a) comprises the following steps:

[0044] a-1) The driving process of a vehicle involves many complex factors, such as the current vehicle speed, load, ambient temperature, and remaining battery power; these factors directly affect the endurance of electric vehicles. Therefore, in order to achieve more accurate and efficient remaining mileage prediction, the data of electric vehicles is collected once a minute. The data includes: vehicle speed, vehicle load current, ambient temperature, remaining battery capacity, and current vehicle mileage. a-2) The feature data of the electric vehicle collected every 30 times constitutes a set of input feature data, X i =[V i ,I i ,T i ,C i ],V i is the speed of the car with the i-th set of input feature data, V i =[v1,v2,...,v i ,...,v 30 ],v iis the speed of the car collected at the i-th minute, I i The vehicle load current of the i-th group of input characteristic data, I i =[e1,e2,...,e i ,...,e 30 ],e i is the vehicle load current collected at the i-th minute, T i is the ambient temperature of the i-th group of input characteristic data, T i =[t1,t2,...,t i ,...,t 30 ], t i is the ambient temperature collected at the i-th minute, C i is the remaining battery capacity of the i-th set of input feature data, C i =[c1,c2,...,c i ,...,c 30 ], c i is the remaining battery capacity collected at the i-th minute.

[0045] a-3) Record the final mileage of the vehicle when the battery is exhausted, and the current mileage of the vehicle collected at the i-th minute a i , subtract the current cruising range of the vehicle collected at the i-th minute from the final cruising range mileage a i Get the remaining mileage y in the i-th minute i , Y i =[y1,y2,...,y i ,...,y 30 ].

[0046] In one embodiment of the present invention, step b) comprises the following steps:

[0047] b-1) The feature extraction module is a bidirectional recurrent neural network BRNN.

[0048] b-2) Input the feature data X of the i-th group i The data in the image are input into the two forward RNN layers and convolutional layers of the bidirectional recurrent neural network BRNN in the forward order from time 1 to 30, and the input is the feature X i ′, the i-th group input feature data X i The data in the image are input into the two reverse RNN layers and convolutional layers of the bidirectional recurrent neural network BRNN in reverse order from time 30 to 1, and the output is the feature X. i ″, feature X i ' and characteristic X i ″ is sequentially input into the fully connected layer and Dropout layer of the bidirectional recurrent neural network BRNN, and the final extracted deep feature vector F is outputi .

[0049] In one embodiment of the present invention, step c) comprises the following steps:

[0050] c-1) The multi-reservoir echo state network model DAMESN based on scaled dot product attention mechanism consists of an input layer, an internal reservoir unit, a dot product scaled attention mechanism, and an output layer.

[0051] c-2) Deep feature vector F i The dimension is Q. The input layer of the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism consists of Q neurons. i Each dimension of is input into the corresponding neuron of the input layer, and the output is the feature vector F i ′.

[0052] c-3) The internal reservoir unit of the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism consists of M internal storage layers, each of which consists of P sparsely connected neurons. i ′ is input into M internal storage layers in sequence, and the output is the feature vector F i ″.

[0053] c-4) The dot product scaling attention mechanism of the multi-reservoir echo state network model DAMESN based on the scaling dot product attention mechanism is composed of a linear layer, a first dot product layer, a scaling layer, a softmax activation function, and a second dot product layer. i ″ is sequentially input into the linear layer of the dot product scaling attention mechanism, the first dot product layer for dot product calculation, the scaling layer, the softmax activation function, and the second dot product layer for dot product calculation, and the feature vector F is output. i ″′.

[0054] c-5) The output layer of the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is composed of a softmax activation function, which converts the feature vector F i ″′ is input to the output layer, and the remaining mileage prediction value is output

[0055] In one embodiment of the present invention, step d) comprises the following steps:

[0056] d-1) Through the formula The objective function f is calculated.

[0057] d-2) Using the Bayesian optimization method, the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is trained based on the objective function f to obtain the optimized multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism.

[0058] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the remaining mileage of a new energy vehicle based on a neural network, characterized in that: The steps include: a) Collect N groups of electric vehicle input feature data to obtain an input feature data set X, where X = {X1, X2, ..., X i ,...,X N }, X i For the i-th group of input feature data, i∈{1,...,N}, the remaining mileage corresponding to the i-th group of input feature data is Y i ; b) Establish a feature extraction module and input the feature data X of the i-th group i Input into the feature extraction module, and output the final extracted deep feature vector F i ; c) Establish a multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism, and transform the deep feature vector F i The input is fed into the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism, and the remaining mileage prediction value is output. d) Using the remaining mileage prediction value And the remaining mileage Y i Train the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism to obtain the optimized multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism; e) The optimized multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is deployed into the vehicle control module of the electric vehicle.

2. The method for predicting remaining mileage of new energy vehicles based on neural network according to claim 1, characterized in that: Step a) comprises the following steps: a-1) Collecting data of electric vehicles once a minute, including: vehicle speed, vehicle load current, ambient temperature, remaining battery capacity, and current mileage of the vehicle; a-2) The characteristic data of the electric vehicle collected every 30 times constitutes a set of input characteristic data, X i =[V i ,I i ,T i ,C i ],V i is the speed of the car with the i-th set of input feature data, V i =[v1,v2,...,v i ,...,v 30 ],v i is the speed of the car collected at the i-th minute, I i The vehicle load current of the i-th group of input characteristic data, I i =[e1,e2,...,e i ,...,e 30 ],e i is the vehicle load current collected at the i-th minute, T i is the ambient temperature of the i-th group of input characteristic data, T i =[t1,t2,...,t i ,...,t 30 ], t i is the ambient temperature collected at the i-th minute, C i is the remaining battery capacity of the i-th set of input feature data, C i =[c1,c2,...,c i ,...,c 30 ], c i is the remaining battery capacity collected at the i-th minute; a-3) Record the final mileage of the vehicle when the battery is exhausted, and the current mileage of the vehicle collected at the i-th minute a i , subtract the current cruising range of the vehicle collected at the i-th minute from the final cruising range mileage a i Get the remaining mileage y in the i-th minute i , Y i =[y1,y2,...,y i ,...,y 30 ].

3. The method for predicting remaining mileage of new energy vehicles based on neural network according to claim 1, characterized in that: Step b) comprises the following steps: b-1) The feature extraction module is a bidirectional recurrent neural network BRNN; b-2) Input the feature data X of the i-th group i The data in the image are input into the two forward RNN layers and convolutional layers of the bidirectional recurrent neural network BRNN in the forward order from time 1 to 30, and the input is the feature X i ′, the i-th group input feature data X i The data in the image are input into the two reverse RNN layers and convolutional layers of the bidirectional recurrent neural network BRNN in reverse order from time 30 to 1, and the output is the feature X. i ″, feature X i ' and characteristic X i ″ is sequentially input into the fully connected layer and Dropout layer of the bidirectional recurrent neural network BRNN, and the final extracted deep feature vector F is output i .

4. The method for predicting remaining mileage of new energy vehicles based on neural network according to claim 1, characterized in that: Step c) comprises the following steps: c-1) The multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism consists of an input layer, an internal reservoir unit, a dot product scaled attention mechanism, and an output layer; c-2) Deep feature vector F i The dimension is Q. The input layer of the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism consists of Q neurons. i Each dimension of is input into the corresponding neuron of the input layer, and the output is the feature vector F i ′; c-3) The internal reservoir unit of the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism consists of M internal storage layers, each of which consists of P sparsely connected neurons. i ′ is input into M internal storage layers in sequence, and the output is the feature vector F i ″; c-4) The dot product scaling attention mechanism of the multi-reservoir echo state network model DAMESN based on the scaling dot product attention mechanism is composed of a linear layer, a first dot product layer, a scaling layer, a softmax activation function, and a second dot product layer. i ″ is sequentially input into the linear layer of the dot product scaling attention mechanism, the first dot product layer for dot product calculation, the scaling layer, the softmax activation function, and the second dot product layer for dot product calculation, and the feature vector F is output. i ″′; c-5) The output layer of the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is composed of a softmax activation function, which converts the feature vector F i ″′ is input to the output layer, and the remaining mileage prediction value is output 5. The method for predicting remaining mileage of new energy vehicles based on neural network according to claim 1, characterized in that: Step d) comprises the following steps: d-1) Through the formula Calculate the objective function f; d-2) Using the Bayesian optimization method, the multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism is trained based on the objective function f to obtain the optimized multi-reservoir echo state network model DAMESN based on the scaled dot product attention mechanism.

Images