Electric vehicle travel energy consumption rate prediction method based on switching type deep learning

Abstract

The invention is suitable for the technical field of electric vehicles, and provides an electric vehicle travel energy consumption rate prediction method based on switching type deep learning. A switching mode of a travel starting stage and a driving stage is designed. In the starting stage of the journey, a feedforward neural network model is used for predicting by using historical data of a target vehicle, and when the model collects enough driving data, the model is automatically switched to a time sequence prediction model for predicting the real-time energy consumption rate. The design not only can flexibly cope with the driving style and the external environment change, but also can obviously improve the prediction accuracy. Meanwhile, factors such as the vehicle speed, the start-stop SOC, the temperature, the idle speed ratio and the idle speed energy consumption rate are considered, compared with an existing method, input characteristics of more dimensions are included, and the prediction accuracy of the energy consumption rate of the electric vehicle is improved. In addition, the energy consumption rate prediction model provided by the invention can more accurately capture subtle differences of energy consumption rate changes, and the prediction precision of the energy consumption rate of the electric vehicle is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric vehicles, and in particular relates to a method for predicting the travel energy consumption rate of an electric vehicle based on switching deep learning. Background Art

[0002] In recent years, the rapid development of electric vehicles has benefited from the support of national policies, the growth of market demand and the promotion of technological innovation. The electrification of transportation has been widely regarded as a necessary means to improve energy efficiency and reduce emissions. By the first half of 2024, the proportion of new energy vehicles in China has exceeded 50%, and electric vehicles are booming.

[0003] The energy consumption rate of electric vehicles, referred to as the Energy Consumption Rate (ECR), is the amount of electrical energy consumed by electric vehicles per kilometer of driving, measured in watts per kilometer (Wh / km). It is a key indicator for measuring the energy efficiency of pure electric vehicles and has an important impact on the energy efficiency, environment and economic benefits of the transportation system. Therefore, accurate prediction of the energy consumption rate of electric vehicles has become a focus of attention for the industry and the public. However, the uncertainty of energy consumption rate prediction, especially in cold areas and complex driving conditions, has caused mileage anxiety, which has become one of the main factors hindering consumers from buying electric vehicles. Accurate real-time energy consumption prediction is crucial to the popularization of electric vehicles. The current methods for predicting the energy consumption rate of electric vehicles have the following main problems: First, traditional prediction methods mainly rely on historical data of target vehicles, but ignore the changes in real-time vehicle status, traffic flow, road conditions and environmental conditions. Under complex and changing road conditions and climate conditions, this static model is difficult to accurately reflect the dynamic changes in actual driving, resulting in deviations in energy consumption prediction results.

[0004] Secondly, data-driven energy consumption rate prediction models usually consider fewer influencing factors, while the energy consumption rate of a vehicle is affected by multiple factors, such as vehicle speed, temperature, idle ratio, etc. Therefore, the model used by the traditional prediction method cannot capture the complex interactions between more factors, resulting in low accuracy in the prediction of the energy consumption rate of electric vehicles.

[0005] Furthermore, mechanism-based models usually simplify complex actual driving environments and operating conditions, and rely on preset physical parameters and formulas, making it impossible to make personalized predictions, further reducing the accuracy of energy consumption rate predictions for electric vehicles.

[0006] In response to the above problems, the present invention proposes a method for predicting the travel energy consumption rate of electric vehicles based on switching deep learning. Summary of the invention

[0007] The purpose of the present invention is to provide a method for predicting the travel energy consumption rate of an electric vehicle based on switching deep learning, aiming to solve the problems raised in the above-mentioned background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions: A method for predicting electric vehicle travel energy consumption rate based on switching deep learning, comprising the following steps: Step S1, data collection: collecting various data related to vehicle driving, including vehicle speed, temperature, idle speed ratio, idle speed energy consumption rate and start-stop SOC; Step S2, pre-switching prediction: at the beginning of the trip, a feedforward neural network model is used to predict the energy consumption rate using the historical data of the target vehicle; Step S3, prediction after switching: During the driving phase, a time series prediction model is used to predict the energy consumption rate using the time series information of the target vehicle at the beginning of the journey and collected in real time.

[0009] Furthermore, the feedforward neural network model adopts an MLP model, and the MLP model includes an input layer, a hidden layer, and an output layer; The input layer receives input features, assuming that the input features are vectors: X=[x1,x2,…,x n ], where X is the entire input feature vector, including all input features; x1,x2,…,x n For each element in the input feature vector, that is, each feature or variable; n is the number of features; the input layer passes the feature vector to the first hidden layer; The neurons in the hidden layer receive the input features of the first input layer and perform linear transformation through weighted summation; the output values ​​of the linear transformation are nonlinearly transformed through a nonlinear activation function; The output layer outputs the predicted value of the electric vehicle energy consumption rate, and the calculation formula is: ; in, is the output value, i.e., the predicted value of the energy consumption rate of the electric vehicle; It is L The weight matrix of the layer, which combines the previous layer ( L- 1 layer) is converted into the weighted input of the current layer; It is L- The activation value vector of layer 1, which is the output of the previous layer; It is L The bias vector of the layer, which adds a constant offset to each node of the current layer; L is the total number of layers, including hidden layers and output layers.

[0010] Furthermore, the specific method of performing the linear transformation is as follows: For l The neurons in the layer i The input of a neuron is the weighted sum of the outputs of all neurons in the previous layer, and the formula is: ; in, It is l Tier i The input of a neuron, It is l Tier i The neurons in the previous layer j The weights between neurons, It is the previous level j The activation value of a neuron, No. l Tier i The bias of a neuron, n l-1 is the number of neurons in the previous layer; The output value of the linear transformation is transformed nonlinearly through a nonlinear activation function: ; in, Indicates L- The activation value vector of layer 1; represents the activation function; express L- The output value of the 1-layer linear transformation.

[0011] Furthermore, the time series prediction model adopts an LSTM model, and the LSTM model includes an input layer, an LSTM layer, a fully connected layer and an output layer; In the LSTM layer, the LSTM layer consists of multiple LSTM units, each of which is responsible for processing the input data of one time step; the LSTM unit contains an input gate, a forget gate, a candidate unit state, and an output gate; Forget Gate f t The calculation formula is: ; in, is the activation function; is the input weight matrix corresponding to the forget gate; is the time step t The input vector of is the recursive weight matrix corresponding to the forget gate; is the previous time step t- A hidden state vector of 1; is the bias vector of the forget gate; Input Gate i t The calculation formula is: ; in, is the input weight matrix corresponding to the input gate; is the recursive weight matrix corresponding to the input gate; is the bias vector of the input gate; Candidate unit status The calculation formula is: ; in, is the input weight matrix corresponding to the candidate memory cell state; is the recursive weight matrix of the candidate memory cell state; is the bias vector of the candidate memory cell state; Unit Status S t The calculation formula is: ; in, is the time step t- 1's unit status; Output Gate o t The calculation formula is: ; in, is the input weight matrix corresponding to the output gate; is the recursive weight matrix corresponding to the output gate; is the bias vector of the output gate; Hidden State h t is the output of the LSTM unit at each time step, passed to the next layer or the next time step, and the calculation formula is: .

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. Compared with the traditional electric vehicle energy consumption rate prediction model, the present invention designs a switching mode between the start of the journey and the driving stage. At the start of the journey, the feedforward neural network model is used to make predictions using the historical data of the target vehicle. When the model collects enough driving data, it will automatically switch to the time series prediction model to predict the real-time energy consumption rate. This design enables the model to flexibly respond to the driver's driving style and the dynamic changes of the external environment, thereby improving the accuracy of the electric vehicle energy consumption rate prediction.

[0013] 2. Compared with the traditional data-driven electric vehicle energy consumption rate prediction model, the feedforward neural network model and time series prediction model adopted by the present invention can receive multi-dimensional data as input, and fully consider multiple key influencing factors such as vehicle speed, start-stop SOC, temperature, idle speed ratio, idle speed energy consumption rate, etc. Compared with the existing methods, the present invention incorporates more dimensional input features and improves the accuracy of electric vehicle energy consumption rate prediction.

[0014] 3. The energy consumption rate prediction model proposed in the present invention can perform personalized prediction analysis through the collected time series data. This feature enables the model to more accurately capture the subtle differences in energy consumption rate changes, thereby effectively improving the accuracy of electric vehicle energy consumption rate prediction. It solves the problem that traditional mechanism models are difficult to accurately describe when faced with the dynamic and nonlinear characteristics of various factors in the energy consumption rate prediction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the present invention.

[0016] Figure 2 It is the box diagram of energy consumption in different speed ranges.

[0017] Figure 3 This is an analysis chart of idle energy consumption rate and idle ratio in different regions.

[0018] Figure 4 Switching architecture diagrams for the start of the journey phase / driving phase.

[0019] Figure 5 This is a comparison chart of the predicted and actual values ​​of the energy consumption rate of electric vehicles. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0021] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0022] like Figure 1 As shown, it is a flowchart of a method for predicting the energy consumption rate of an electric vehicle based on switching deep learning provided by an embodiment of the present invention. The method predicts the energy consumption rate of an electric vehicle based on an electric vehicle energy consumption rate prediction model. The model is implemented by a software system, running on an Ubuntu 18.04 operating system, and uses the PyTorch2.3.1 package to construct a feedforward neural network model before switching and a timing prediction model after switching.

[0023] The method comprises the following steps: Data collection: Collect various data related to vehicle driving, including vehicle speed, temperature (minimum temperature, maximum temperature), idle ratio, idle energy consumption rate, and start and end SOC (starting SOC, ending SOC), to provide a basis for subsequent energy consumption rate model prediction.

[0024] Model building, training preparation and data set division: Build the model of the entire algorithm in PyCharm2024.1 software, select the appropriate optimizer, learning rate, training cycle, batch size and time step, etc. Then, divide the data set into training set, validation set and test set. The training set is used to train the parameters of the energy consumption rate prediction model, and the validation set is used to adjust the hyperparameters during the training process.

[0025] Prediction before switching: At the beginning of the journey, a feedforward neural network model is used to predict the energy consumption rate using the historical data of the target vehicle. The feedforward neural network model uses a multi-layer perceptron (MLP) model as an example.

[0026] Prediction after switching: During the driving phase, the time series prediction model is used to predict the energy consumption rate using the time series information collected at the beginning of the trip and in real time. The time series prediction model uses the Long Short-Term Memory (LSTM) model as an example.

[0027] Model training and verification: Use the training set to train the energy consumption rate prediction model. Through model training, accurate prediction of the energy consumption rate of electric vehicles can be achieved. During the training process, the validation set is used to adjust the model's hyperparameters to achieve the best prediction effect. After the training is completed, the test set is used to verify the actual prediction effect of the model to ensure the accuracy and reliability of the model.

[0028] 1) Input feature selection for electric vehicle energy consumption rate prediction in the data collection step; The vehicle energy consumption rate consists of the driving segment energy consumption rate and the idling energy consumption rate. All data are divided into the idling segment and the driving segment. The idling segment is defined as the state where the vehicle speed is 0, otherwise it is defined as the driving segment. The energy consumption of each driving segment can be calculated by integrating the product of the battery pack current and voltage: ; in, E is the energy consumed in the drive segment, in watt-hours (Wh); U is the total voltage of the battery pack in volts (V); I is the total current of the battery pack in amperes (A); dt Represents tiny increments of time.

[0029] The mileage is obtained by subtracting the initial mileage from the end mileage of the driving segment: ; in, is the driving segment mileage, is the distance traveled at the end of the driving section, is the distance traveled at the beginning of the driving segment.

[0030] Energy consumption rate of driving section: ; in, is the energy consumption rate per unit distance, in watt-hours per kilometer (Wh / km); It is the energy consumption of a driving section, measured in watt-hours (Wh).

[0031] Therefore, based on the above mechanism knowledge, the energy consumption rate of each driving segment of the electric vehicle data can be analyzed to obtain the following Figure 2 The average vehicle speed shown is the average vehicle speed of each driving segment, and the vertical axis is a box plot of the average energy consumption rate in this speed range. The results show that the average vehicle speed has the greatest impact on the energy consumption rate of actual road driving. The energy consumption rate of the driving segment drops sharply as the speed increases, and then remains stable (average vehicle speed is above 25 km / h).

[0032] Under low average speed conditions, the main reason for the high energy consumption rate of the drive section is the low speed and torque, which leads to low motor efficiency. In addition, frequent starts and stops also lead to high energy consumption rates. According to the increase in average speed, the reduction in energy consumption rate is due to the improvement of motor operating efficiency and driving efficiency (mileage in a fixed time period). Under high speed conditions, the slight increase in energy consumption rate is due to the increase in air resistance and rolling resistance.

[0033] The energy consumption rate at idle speed (in watt-hours per hour, Wh / h) is calculated as follows: ; in, It is the energy consumption rate in the idle speed range.

[0034] The energy consumption rate in the idle section is affected by the idle ratio. The idle ratios in different cities vary significantly, which affects energy consumption. The analysis of idle energy consumption rate and idle ratio in different regions is shown in the figure below. Figure 3 shown.

[0035] Therefore, based on the above analysis and combined with the knowledge of energy consumption rate mechanism, we finally screened out the characteristic variables that can highly map the energy consumption rate characteristics of electric vehicles: vehicle speed, idle ratio, idle energy consumption rate, start-stop SOC and temperature.

[0036] 2) The start of the journey; The beginning of a journey usually refers to the model used when the initial data is limited and the time series data is incomplete or insufficient. At this time, the MLP model is a common choice because of its simple structure and low data requirements. The MLP model consists of an input layer, a hidden layer, and an output layer. Each neuron in each layer is connected to all neurons in the previous layer to form a fully connected network structure. Through nonlinear activation functions, the MLP model can perform nonlinear transformations on input features to adapt to more complex data distributions.

[0037] The input layer receives static input features of different dimensions (such as vehicle speed, temperature, etc.), which can be extracted from historical data. Let the input feature be a vector: X=[x1,x2,…,x n ], where X is the entire input feature vector, including all input features; x1,x2,…,x n is each element in the input feature vector, that is, each feature or variable; n is the number of features. The input layer passes the feature vector to the first hidden layer. The input here is processed data, that is, data that has been preprocessed by standardization or normalization.

[0038] The hidden layer is the core part of MLP, which performs linear transformation and activation function. The linear transformation is that the neurons in the hidden layer receive the input features of the first layer of input layer and perform linear transformation through weighted summation. The specific way to perform linear transformation is as follows: For l The neurons in the layer i The input of a neuron is the weighted sum of the outputs of all neurons in the previous layer, and the formula is: ; in, It is l Tier i The input of a neuron, It is l Tier i The neurons in the previous layer j The weights between neurons, It is the previous level j The activation value of a neuron, No. l Tier i The bias of a neuron, n l-1 is the number of neurons in the previous layer.

[0039] The output value of the linear transformation is transformed nonlinearly through a nonlinear activation function: ; in, IndicatesL- The activation value vector of layer 1; represents the activation function; express L- The output value of the 1-layer linear transformation.

[0040] In the output layer, the predicted value of the electric vehicle energy consumption rate is output, and the calculation formula is: ; in, is the output value, i.e., the predicted value of the energy consumption rate of the electric vehicle; It is L The weight matrix of the layer, which combines the previous layer ( L- 1 layer) is converted into the weighted input of the current layer; It is L- The activation value vector of layer 1, which is the output of the previous layer; It is L The bias vector of the layer, which adds a constant offset to each node of the current layer; L is the total number of layers, including hidden layers and output layers.

[0041] 3) LSTM model in the driving phase; Since the amount of driving data at the beginning of the journey is small, the MLP model is first used to use historical data as input. When the model collects enough driving data, it can make better predictions based on real-time driving data. However, the MLP model is not good at processing time series data, so the LSTM model is used as the prediction model for the driving stage. The LSTM model is good at extracting long-term dependencies in the data, which helps to better predict the energy consumption rate of electric vehicles in real time. The LSTM model includes an input layer, an LSTM layer, a fully connected layer, and an output layer, which enables the module to learn and remember long-term dependency features.

[0042] In the input layer, the input layer receives real-time driving data, such as vehicle speed sequence, idle energy consumption rate, idle ratio and other input features, and passes them to the LSTM layer.

[0043] In the LSTM layer, the LSTM layer consists of multiple LSTM units, each of which is responsible for processing the input data of one time step. The LSTM unit contains four main components: input gate, forget gate, candidate unit state and output gate.

[0044] Forget Gate Controls how much old information is discarded, calculated as: ; in, is the activation function; is the input weight matrix corresponding to the forget gate; is the time step t The input vector of is the recursive weight matrix corresponding to the forget gate; is the previous time step t- A hidden state vector of 1; is the bias vector of the forget gate; Input Gate Determines how much new information is introduced into the memory unit, and the calculation formula is: ; in, is the input weight matrix corresponding to the input gate; is the recursive weight matrix corresponding to the input gate; is the bias vector of the input gate; Candidate unit status The calculation formula is: ; in, is the input weight matrix corresponding to the candidate memory cell state; is the recursive weight matrix of the candidate memory cell state; is the bias vector of the candidate memory cell state; Unit Status , is updated at each time step, which consists of two parts: one is the result of multiplying the previous memory state by the forget gate, and the other is the candidate unit state AND input gate The result of multiplication is calculated as follows: ; in, is the time step t- 1's unit status; Output Gate Determines how much information to output from the memory cell to the next time step or the output layer of the network. The calculation formula is: ; in, is the input weight matrix corresponding to the output gate; is the recursive weight matrix corresponding to the output gate; is the bias vector of the output gate; Hidden State , is the output of the LSTM unit at each time step, passed to the next layer or the next time step, and the calculation formula is: ; In the fully connected layer, the output of the LSTM layer is converted into a fixed-size vector. This vector contains the features extracted from the input data, and the features extracted by the LSTM layer are further processed to prepare for the prediction of the output layer.

[0045] In the output layer, the final prediction result, i.e., the energy consumption rate of the electric vehicle, is generated.

[0046] Switching principle; Since the trip energy consumption rate and the timing information of the driving process are strongly correlated, the present invention selects a time series model as a prediction model during the driving process, and it requires a period of driving data as model input. Therefore, at the beginning of the journey, since the model cannot collect enough driving data for timing prediction, the feedforward neural network model is first used to use the historical data of the target vehicle (including vehicle location and time information, average vehicle speed, idle energy consumption rate, and idle ratio) to predict the energy consumption rate; when the model collects enough timing data during driving (for example, the electric vehicle can be set to switch after driving for 25 minutes), it will enter the driving stage and will automatically switch to the timing prediction model, using the timing data of the current driving process (real-time vehicle speed sequence, real-time idle energy consumption rate, real-time idle ratio, temperature) to predict the real-time energy consumption rate. The trip start stage / driving stage switching architecture is as follows: Figure 4 shown.

[0047] Example 1: Training and verification of an electric vehicle energy consumption rate prediction model; The electric vehicle driving data collected in real scenarios is divided into training set, validation set and test set in chronological order. Specifically, the first 90% of the data is used to train model parameters and adjust hyperparameters, and the last 10% of the data is used for the test set to evaluate the performance of the model. The entire model is built and trained based on the PyTorch framework and runs on a PC server equipped with an Intel(R) Core(TM) i7-13700KF 3.40 GHz processor and a 12 GB NVIDIA GeForce RTX 4070 Ti GPU.

[0048] In order to prevent the overfitting problem of the deep learning model, the model design adopts a Dropout strategy with a probability of 0.3, which is applied between the convolutional layer, the recurrent layer and the fully connected layer. The optimizer uses the Adam algorithm, and the loss function is the mean square error (MSE). By adjusting the hyperparameters on the validation set, the parameters determined by the final model include a time step of 25, a batch size of 64, and a bidirectional LSTM containing 64 hidden neurons to capture time series features. The model training was carried out for 30 cycles, and the evaluation indicators were the mean absolute percentage error (MAPE) and the coefficient of determination ( R 2) in order to quantitatively analyze the experimental results.

[0049] ; ; in, n is the sample size, y i is the actual remaining energy consumption rate of electric vehicles, is the residual value of energy consumption rate predicted by the electric vehicle energy consumption rate prediction model, It is the average value of the actual energy consumption rate. MAPE is a statistical indicator used to measure the accuracy of the prediction model. It is mainly used to evaluate the deviation between the predicted value and the actual value of the model. MAPE represents the average percentage of prediction error and can intuitively reflect the accuracy of the prediction result relative to the actual result. The smaller its value, the higher the prediction accuracy of the model. R 2 A measure of how well the variability predicted by the model matches the actual variability in the data. R 2 The value of is usually between 0 and 1. The closer the value is to 1, the better the model explains the variation of the data.

[0050] The trained model is used to perform inference analysis on the test set, and the comparison between the predicted and actual values ​​of the electric vehicle energy consumption rate is obtained. Figure 5 As shown in the figure, the predicted value is basically consistent with the actual value, and the energy consumption rate prediction model is further calculated. R 2 and MAPE are 0.99 and 0.01 respectively, indicating that the model has very high prediction accuracy and good data variability explanation ability. The above results fully demonstrate the effectiveness and versatility of the energy consumption rate prediction model in practical applications. It is worth mentioning that the model is mainly based on data-driven and mechanism knowledge. Compared with the model that only relies on a single mechanism or a single data-driven, it has higher accuracy and can accurately estimate the energy consumption rate of actual electric vehicles in the full range of operating conditions. Therefore, it can be confirmed that the electric vehicle travel energy consumption rate prediction method based on switching deep learning proposed in the present invention can realize real-time and accurate prediction of the energy consumption rate of electric vehicles.

[0051] The above are only preferred embodiments of the present invention. It should be pointed out that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention. These should also be regarded as the protection scope of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A method for predicting electric vehicle travel energy consumption rate based on switching deep learning, characterized in that: The following steps are involved: Step S1, data collection: collecting various data related to vehicle driving, including vehicle speed, temperature, idle speed ratio, idle speed energy consumption rate and start-stop SOC; Step S2, pre-switching prediction: at the beginning of the trip, a feedforward neural network model is used to predict the energy consumption rate using the historical data of the target vehicle; Step S3, prediction after switching: During the driving phase, a time series prediction model is used to predict the energy consumption rate using the time series information of the target vehicle at the beginning of the journey and collected in real time.

2. The electric vehicle travel energy consumption rate prediction method based on switching deep learning according to claim 1 is characterized in that: The feedforward neural network model adopts an MLP model, and the MLP model includes an input layer, a hidden layer and an output layer; The input layer receives input features, assuming that the input features are vectors: X=[x1,x2,…,x n ], where X is the entire input feature vector, including all input features; x1,x2,…,x n For each element in the input feature vector, that is, each feature or variable; n is the number of features; the input layer passes the feature vector to the first hidden layer; The neurons in the hidden layer receive the input features of the first input layer and perform linear transformation through weighted summation; the output values ​​of the linear transformation are nonlinearly transformed through a nonlinear activation function; The output layer outputs the predicted value of the electric vehicle energy consumption rate, and the calculation formula is: ； in, is the output value, i.e., the predicted value of the energy consumption rate of the electric vehicle; It is L The weight matrix of the layer; It is L- The activation value vector of layer 1, which is the output of the previous layer; It is L The bias vector of the layer; L is the total number of layers, including hidden layers and output layers.

3. The electric vehicle travel energy consumption rate prediction method based on switching deep learning according to claim 2 is characterized in that: The specific method of performing the linear transformation is as follows: For l The neurons in the layer i The input of a neuron is the weighted sum of the outputs of all neurons in the previous layer, and the formula is: ； in, It is l Tier i The input of a neuron, It is l Tier i The neurons in the previous layer j The weights between neurons, It is the previous level j The activation value of a neuron, No. l Tier i The bias of a neuron, n l-1 is the number of neurons in the previous layer; The output value of the linear transformation is transformed nonlinearly through a nonlinear activation function: ； in, Indicates L- The activation value vector of layer 1; represents the activation function; express L- The output value of the 1-layer linear transformation.

4. The electric vehicle travel energy consumption rate prediction method based on switching deep learning according to claim 1 is characterized in that: The time series prediction model adopts an LSTM model, which includes an input layer, an LSTM layer, a fully connected layer and an output layer; In the LSTM layer, the LSTM layer consists of multiple LSTM units, each of which is responsible for processing the input data of one time step; the LSTM unit contains an input gate, a forget gate, a candidate unit state, and an output gate; Forget Gate f t The calculation formula is: ； in, is the activation function; is the input weight matrix corresponding to the forget gate; is the time step t The input vector of is the recursive weight matrix corresponding to the forget gate; is the previous time step t- A hidden state vector of 1; is the bias vector of the forget gate; Input Gate i t The calculation formula is: ； in, is the input weight matrix corresponding to the input gate; is the recursive weight matrix corresponding to the input gate; is the bias vector of the input gate; Candidate unit status The calculation formula is: ； in, is the input weight matrix corresponding to the candidate memory cell state; is the recursive weight matrix of the candidate memory cell state; is the bias vector of the candidate memory cell state; Unit Status S t The calculation formula is: ； in, is the time step t- 1's unit status; Output Gate o t The calculation formula is: ； in, is the input weight matrix corresponding to the output gate; is the recursive weight matrix corresponding to the output gate; is the bias vector of the output gate; Hidden State h t is the output of the LSTM unit at each time step, passed to the next layer or the next time step, and the calculation formula is: 。

Images