WO2022077693A1 - Load prediction model training method and apparatus, storage medium, and device - Google Patents

Abstract

Disclosed in the present invention are a load prediction model training method and apparatus for use in an electric vehicle charging station, a storage medium, and a device. The training method comprises: obtaining the historical load state data of a charging station at a plurality of moments, wherein the historical load state data at each moment comprises a plurality of types of historical variable data and corresponding real load data; sequentially and independently predicting each type of historical variable data at each moment by using a preset model, so as to generate a plurality of groups of predicted load data; and according to a reinforcement learning method, using the plurality of groups of predicted load data and the real load data to train the weight data group of a load prediction model to be trained. The prediction precision is improved by using a plurality of types of historical variable data. Moreover, extracting more useful information from inputted data by using a phase space reconstruction technology improves the calculation performance of the model. Using a GRU neural network to predict the inputted data accelerates a calculation speed, and improves the prediction precision of the model by combining the Q learning algorithm.

Description

Training method and training device, storage medium and equipment for load prediction model technical field

The invention belongs to the technical field of battery management, and in particular, relates to a training method and a prediction device for a load prediction model of an electric vehicle charging station, a computer-readable storage medium, and computer equipment.

Background technique

Energy shortages and climate change are major challenges the world is facing. The exhaust gas emitted by internal combustion engine vehicles powered by traditional fossil fuels not only causes great air pollution, but also further aggravates the greenhouse effect. Countries around the world urgently need to take effective response measures and policies. Vigorously developing electric vehicles can effectively solve these problems. An electric vehicle is a zero-emission system that is powered by electricity and does not produce any substances that pollute the environment during driving. Electric vehicles can also use a variety of energy sources, including wind energy, solar energy and other clean energy sources, thus greatly reducing the consumption of petrochemical resources. At the same time, the powertrain and electric motor of electric vehicles are more efficient and environmentally friendly than traditional combustion engine vehicles.

The development of the electric vehicle industry is the general trend, which will inevitably bring new changes and new problems to the electric energy field. As the number of electric vehicles increases, the demand for charging stations will naturally increase. The charging station will face the situation of continuous uninterrupted use, and the power system will also face the problem of excessive load. In order to ensure the charging demand of customers and the stable operation of the power system, the load forecasting of electric vehicle charging stations is extremely necessary and critical. Accurate load forecasting can help power companies reasonably allocate and develop power infrastructure and load configuration.

The existing traditional technical solutions such as the autoregressive moving average model and the autoregressive integral moving average model do not have the ability to extract nonlinear features. However, the shallow neural network has the problems of model overfitting, sensitivity to random initialization weights, and easy convergence to the local optimum. Therefore, the prediction results are often inaccurate and the error is too large.

Deep reinforcement learning is a subfield of machine learning, a method that combines deep learning with reinforcement learning, and has a wide range of applications. Deep reinforcement learning combines the ability of deep learning to extract hidden features of nonlinear data and the decision-making ability of reinforcement learning, and has the advantages of both.

The problems existing in traditional load forecasting algorithms are generally poor forecasting performance, poor robustness and adaptability. Electric vehicle charging station load has a strong correlation with external factors, and existing methods usually only consider a single external factor, which leads to low prediction accuracy.

SUMMARY OF THE INVENTION

(1) Technical problem to be solved by the present invention

The technical problem solved by the present invention is: how to consider multiple external factors in the model training process to improve the model prediction accuracy.

(2) Technical scheme adopted in the present invention

A training method for a load prediction model of an electric vehicle charging station, the training method comprising:

Acquire historical load status data of the charging station at several moments, and the historical load status data at each moment includes multiple types of historical variable data and corresponding real load data;

Use the preset model to make separate predictions for each type of historical variable data at each moment in turn to generate multiple sets of forecasted load data;

According to the reinforcement learning method, the weight data sets of the load prediction model to be trained are trained by using multiple sets of the predicted load data and the real load data.

Preferably, after acquiring the historical load state data of the charging station, the training method further includes:

Phase space reconstruction is performed on the multiple types of the historical load state data to generate multiple types of reconstructed variable data and reconstructed real load data.

Preferably, the method of using a preset model to sequentially predict each type of historical variable data at each moment individually to generate multiple sets of predicted load data includes:

Each type of reconstructed variable data at each moment is input into the cyclic gate unit network model in turn, and the cyclic gate unit network model outputs multiple sets of predicted load data.

Preferably, the reinforcement learning method is a Q-learning method, and the method for training the weight data groups of the load prediction model to be trained by using multiple sets of the predicted load data and the real load data includes:

Constructing and initializing a state matrix and an action matrix, wherein the state matrix is composed of the weight data group to be trained, and the action matrix is composed of the variation of the weight;

The state matrix executes the action matrix according to the preset strategy to update the weight data set;

Calculate a loss function according to the updated weight data set, multiple sets of predicted load data and the reconstructed real load data, and calculate a reward factor according to the loss function;

Update the state matrix and the action matrix according to the reward factor;

Repeat the above steps until the iteration conditions are met.

Preferably, the calculation formula of the loss function is:

in,

w ₁ , w ₂ ... w _i are weight data groups to be trained, X ₁ , X ₂ ... X _i are multi-class historical variable data, f(X ₁ ), f(X ₂ )...f(X _i ) are multiple Group predicted load data, Y is the real load data, N is the number of groups of historical variable data.

The invention also discloses a training device for a load prediction model of an electric vehicle charging station, the training device comprising:

an acquisition module, used for acquiring historical load status data of the charging station at several moments, the historical load status data at each moment including multiple types of historical variable data and corresponding real load data;

The forecasting module is used to make separate forecasts for each type of historical variable data at each moment in turn to generate multiple sets of forecasted load data;

The training module is used for training the weight data groups of the load prediction model to be trained according to the reinforcement learning method and using multiple sets of the predicted load data and the real load data.

Preferably, the training device further comprises:

The data reconstruction module is used for performing phase space reconstruction on the multiple types of the historical load state data to generate multiple types of reconstructed variable data and reconstructed real load data.

Preferably, the training module includes:

an initialization unit, configured to construct and initialize a state matrix and an action matrix, wherein the state matrix is composed of the weight data group to be trained, and the action matrix is composed of the variation of the weight;

an execution unit, used to make the state matrix execute the action matrix according to the preset strategy to update the weight data group;

a calculation unit, configured to calculate a loss function according to the updated weight data group, multiple groups of predicted load data and the reconstructed real load data, and calculate a reward factor according to the loss function;

an update unit, configured to update the state matrix and the action matrix according to the reward factor.

The invention also discloses a computer-readable storage medium storing a training program for a load prediction model of an electric vehicle charging station, and the training program for a load prediction model of an electric vehicle charging station When executed by a processor, the above-described training method for a load prediction model for an electric vehicle charging station is implemented.

The invention also discloses a computer device, the computer device comprises a computer-readable storage medium, a processor and a training program for a load prediction model of an electric vehicle charging station stored in the computer-readable storage medium, When the training program for the load prediction model for the electric vehicle charging station is executed by the processor, the above-mentioned training method for the load prediction model for the electric vehicle charging station is implemented.

(3) Beneficial effects

The training method for the load prediction model of the electric vehicle charging station disclosed in the present invention solves the problem that the prediction result is inaccurate because some methods do not consider various factors by using various historical variable data. At the same time, the phase space reconstruction technique is used to extract more useful information from the input data, which improves the computational performance of the model. The GRU neural network is also used to predict the input data, which solves the problems of traditional LSTM with many parameters and slow calculation speed. Finally, the Q-learning algorithm using the ε-greedy strategy is used to train the combined weights of the prediction results of the model, which improves the prediction accuracy of the model.

Description of drawings

1 is a flowchart of a training method for a charging station load prediction model according to Embodiment 1 of the present invention;

FIG. 2 is another flowchart of the training method of the charging station load prediction model according to the first embodiment of the present invention;

3 is a schematic diagram of a GRU model according to Embodiment 1 of the present invention;

4 is a Q-learning training flow chart of Embodiment 1 of the present invention;

5 is a schematic diagram of a training device according to Embodiment 2 of the present invention;

6 is a schematic diagram of a training module according to Embodiment 2 of the present invention;

7 is a schematic diagram of an overall training process of the training device according to the second embodiment of the present invention;

FIG. 8 is a schematic block diagram of a computer device according to an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

Before describing the technical solution of the present application in detail, first briefly describe the inventive concept of the present application: the electric vehicle charging station load has a strong correlation with external factors, and there are various external factors, but the existing method only considers a single external factor This application first obtains multiple types of historical variable data, that is, a variety of external factors, and predicts each historical variable data separately, and then uses the reinforcement learning method to train to obtain the best weight combination. The influence of various variables can improve the prediction accuracy of the model.

Specifically, as shown in FIG. 1 , the training method for a load prediction model of an electric vehicle charging station according to Embodiment 1 includes the following steps:

Step S10: obtaining historical load status data of the charging station at several moments, the historical load status data at each moment including multiple types of historical variable data and corresponding real load data;

Step S20: using the preset model to perform separate prediction on each type of historical variable data at each moment in turn to generate multiple sets of predicted load data;

Step S30: According to the reinforcement learning method, training is performed on the weight data sets of the load prediction model to be trained by using multiple sets of the predicted load data and the real load data.

Exemplarily, in step S10, collect external factor data for 90 consecutive days and collect the charging data of all charging piles in the charging station, that is, external factor data and charging data are taken together as historical variable data, wherein historical variable data includes weather information, External data such as time-of-day information, holiday information, real-time prices and traffic flow, and charging station data including charge volume, charge duration, and charge power. Considering that the charging station can be used for 24 hours, and the randomness of the charging vehicle entering the charging station and the charging start time. The real load data in the first embodiment refers to the superimposed load of all the charging piles of the charging station, that is, the load data of the charging station for a whole day. Select the historical variable data of the previous 90 days as the input X, and the input X can be expanded in the time dimension.

As another embodiment, after step S10, step S11 is further included: preprocessing the historical load state data. Specifically, the historical variable data is divided into a training set and a test set, the training set is used to train model parameters, and the test set is used to test the accuracy of the model; the reason for abnormal data is that due to some interference factors, the data is missing or wrong , the processing method is: for the missing data, if the time interval before and after is less than or equal to the set threshold, the mean value of the data before and after is used to make up; if the time interval before and after is greater than the set threshold, the previous data is used to replace it. The data of the same date type should be used; if a certain type of data is Null, delete the data in the column or complete it with 0; If the variation range of the data is greater than a certain threshold, the average value of the before and after values is taken instead; finally, the input data is normalized.

Among them, x is the historical variable data before normalization, and x' represents the normalized historical variable data.

As another embodiment, after step S11, step S12 is further included: performing phase space reconstruction on the preprocessed historical variable data. Specifically, input multiple sequences X=[x ₁ , x ₂ ,...,x _N ], where N is the number of elements in the input sequence. Input variables can include weather information, time period information, holiday information, real-time price, traffic flow, charging amount, charging time, charging power, etc.

The phase point of the delay sequence is X _t =(x _t ,x _(t+τ) ,x _(t+2τ) ,…,x _[t+(m-1)τ] ) ^T ,t=1,2,…, N-(m-1)τ. The phase space is reconstructed using the delay sequence, and each column in the phase space X has a phase point X _t .

The delay time τ and the embedding dimension m are determined using the C-C method.

Specifically, the associated integral is defined as:

where r is the spatial distance, and ||X _i -X _j || is the Euclidean distance.

The delay time is τ, and the serial correlation of the time series is expressed as:

The local optimal delay time τ ^* is the time at which the first S(m,r,τ)=0 or the minimum value is reached. At the local optimum delay time τ ^* , the reconstructed phase space is under a nearly uniform point distribution.

The correlation interval difference is expressed as:

ΔS(m,τ)=max{S(m,r _j ,τ)}-min{S(m,r _j ,τ)}

When N is greater than 3000, m=2, 3, 4, 5,

j=m-1, s is the standard deviation of the input sequence.

calculate

The global minimum point of τ _w is obtained, and the optimal phase space dimension m ^* can be obtained according to τ _w =(m-1)τ ^* .

Further, the actual load data Y corresponding to the above historical variable data is also subjected to corresponding preprocessing and phase space reconstruction, which will not be repeated here.

Further, in step S20, the cyclic gate unit network model (GRU) is used to independently predict each type of historical variable data, that is, what is the corresponding power station load state data when only one variable is considered. The predicted load data corresponding to each variable is

f is the GRU model function, a total of i historical variable data are input, the GRU model independently predicts each historical variable data, and the prediction result after weight combination is expressed as

w ₁ , w ₂ ... w _i are weight data groups to be trained, X ₁ , X ₂ ... X _i are multi-class historical variable data, f(X ₁ ), f(X ₂ )...f(X _i ) are multiple Group forecast load data.

Specifically, as shown in Figure 3, GRU (Gate Recurrent Unit) is an improved model of LSTM (Long Short Term Memory), which simplifies the structure of LSTM by discarding the storage unit mechanism and introducing update gates to replace input and forget gates. The GRU model consists of an input layer, a hidden layer, and an output layer, where the hidden layer consists of a reset gate r _t and an update gate z _t . Both gates depend on the previous hidden state ht _-1 and the current input _xt . The reset gate _rt determines the new hidden state obtained after using the filtered information and the current input x _t

How many previous hidden states h _t-1 have been filtered before.

The update gate zt controls the previous hidden state _{ht -1} and the next hidden state

, so as to ensure that valid information can flow to the next GRU unit, and each sequence constructed is input into the GRU network to obtain the desired prediction result.

The activation functions in the hidden layer of the GRU network include the sigmoid function and the tanh function, whose expressions are:

The GRU processing formula is as follows:

z _t =σ(U ^z h _t-1 +x _t W ^z )

r _t =σ(U ^r h _t-1 +x _t W ^r )

where h _t is the hidden state

is the new hidden state, W ^z is the updated gate weight matrix, W ^r is the reset gate weight matrix, and W is the hidden state weight matrix.

The specific process of using the GRU model to predict historical variable data is in the prior art, and will not be described in detail here.

Further, as shown in FIG. 4 , in step S30, the reinforcement learning method is a Q learning method, and a method for training the weight data groups of the load prediction model to be trained by using multiple groups of the predicted load data and the real load data. include:

Step S31: Construct and initialize a state matrix and an action matrix, wherein the state matrix S is composed of the weight data set to be trained, and the action matrix a is composed of the variation of the weight. The state matrix S is the horizontal row of the Q table, and the action matrix a is the vertical column of the Q table.

S=[w ₁ ,w ₂ ,..., _wi ]

a=[Δw ₁ ,Δw ₂ ,...,Δw _i ]

Step S32: The state matrix S executes the action matrix a according to the preset strategy to update the weight data set.

As a preferred embodiment, the ε-greedy strategy is used:

ε is a random value in the range (0, 1). Exploration with the probability of ε (Exploration) is to randomly select the action a, and with the probability of 1-ε (Exploitation) to select the action a with the largest Q value.

Step S33: Calculate a loss function according to the updated weight data set, multiple sets of predicted load data and real load data, and calculate a reward factor according to the loss function.

The calculation formula of the loss function L is:

in,

w ₁ , w ₂ ... w _i are weight data groups to be trained, X ₁ , X ₂ ... X _i are multi-class historical variable data, f(X ₁ ), f(X ₂ )...f(X _i ) are multiple Group predicted load data, Y is the real load data, N is the number of groups of historical variable data, that is, N groups

data.

The formula for calculating the reward factor R is as follows:

Step S34: Update the state matrix and the action matrix according to the reward factor;

After the agent performs action a, it calculates the loss function L, calculates and obtains the reward factor R, calculates the Q value and updates the Q table and state S. where α is the learning rate and γ is the discount factor. The larger the learning rate, the less effective it is to retain previous training.

The above steps are repeated until the iterative conditions are met, and the optimal state matrix S can be obtained, thereby obtaining the optimal weight data set.

In practical applications, the test data is input into the trained model for prediction, and the final prediction result is obtained, and then inverse normalization is performed, that is, the load of the charging station in the future is obtained. Among them, the test data also needs to be reconstructed in phase space.

The training method of the load forecasting model disclosed in the first embodiment solves the problem of inaccurate forecasting results caused by some methods failing to consider various factors by using various historical variable data. At the same time, the phase space reconstruction technique is used to extract more useful information from the input data, which improves the computational performance of the model. The GRU neural network is also used to predict the input data, which solves the problems of traditional LSTM with many parameters and slow calculation speed. Finally, the Q-learning algorithm using the ε-greedy strategy is used to train the combined weights of the prediction results of the model, which improves the prediction accuracy of the model.

Embodiment 2

As shown in FIG. 5 , a training device for a load prediction model of an electric vehicle charging station provided in the second embodiment includes an acquisition module 100 , a prediction module 200 , and a training module 300 . The acquisition module 100 is used to acquire historical load status data of the charging station at several moments, and the historical load status data at each moment includes multiple types of historical variable data and corresponding real load data; the prediction module 200 is used for sequentially Each type of historical variable data is independently predicted to generate multiple sets of predicted load data; the training module 300 is used for the weight of the load prediction model to be trained by using multiple sets of the predicted load data and the real load data according to the reinforcement learning method data set for training.

Further, the training device further includes a data reconstruction module 500, and the data reconstruction module 500 is configured to perform phase space reconstruction on the multiple types of the historical variable data to generate multiple types of reconstructed variable data. For the specific process of the phase space reconstruction, refer to the description in Embodiment 1, which is not repeated here.

Further, as shown in FIG. 6 , the training module 400 includes an initialization unit 401 , an execution unit 402 , a calculation unit 403 and an update unit 404 . Wherein, the initialization unit 401 is used to construct and initialize the state matrix and the action matrix, wherein the state matrix is composed of the weight data group to be trained, and the action matrix is composed of the variation of the weight; the execution unit 402 is used to make the state matrix according to The preset strategy execution action matrix is used to update the weight data group; the calculation unit 403 is configured to calculate a loss function according to the updated weight data group, multiple groups of predicted load data and real load data, and calculate the reward factor according to the loss function; update The unit 404 is configured to update the state matrix and the action matrix according to the reward factor. For the specific training process of the training module 400, refer to Embodiment 1, which will not be repeated here.

The overall training process of the training device for the load prediction model of the second embodiment is shown in FIG. 7 .

The present application also discloses a computer-readable storage medium, where the computer-readable storage medium stores a training program for a load prediction model of an electric vehicle charging station, and the training program for a load prediction model of an electric vehicle charging station When executed by a processor, the above-described training method for a load prediction model for an electric vehicle charging station is implemented.

The present application also discloses a computer device. At the hardware level, as shown in FIG. 8 , the terminal includes a processor 12 , an internal bus 13 , a network interface 14 , and a computer-readable storage medium 11 . The processor 12 reads the corresponding computer program from the computer-readable storage medium and then executes it, forming a request processing device on a logical level. Of course, in addition to software implementations, one or more embodiments of this specification do not exclude other implementations, such as logic devices or a combination of software and hardware, etc., that is to say, the execution subjects of the following processing procedures are not limited to each Logic unit, which can also be hardware or logic device. The computer-readable storage medium 11 stores a training program for a load prediction model of an electric vehicle charging station. A training method for a load prediction model of a car charging station.

Computer-readable storage media includes both persistent and non-permanent, removable and non-removable media, and storage of information can be implemented by any method or technology. Information may be computer readable instructions, data structures, modules of programs, or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Flash Memory or other memory technology, Compact Disc Read Only Memory (CD-ROM), Digital Versatile Disc (DVD) or other optical storage , magnetic cassettes, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by computing devices.

The specific embodiments of the present invention have been described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that the principles and spirit of the present invention, which are defined in the scope of the claims and their equivalents, are not departed from. Under the circumstances, these embodiments can be modified and perfected, and these modifications and improvements should also fall within the protection scope of the present invention.

Claims (10)

1. A training method for a load prediction model of an electric vehicle charging station, characterized in that the training method comprises:

   Acquire historical load status data of the charging station at several moments, and the historical load status data at each moment includes multiple types of historical variable data and corresponding real load data;

   Use the preset model to make separate predictions for each type of historical variable data at each moment in turn to generate multiple sets of forecasted load data;

   According to the reinforcement learning method, the weight data sets of the load prediction model to be trained are trained by using multiple sets of the predicted load data and the real load data.
2. The method for training a load prediction model for an electric vehicle charging station according to claim 1, wherein after acquiring the historical load state data of the charging station, the training method further comprises:

   Phase space reconstruction is performed on the multiple types of the historical load state data to generate multiple types of reconstructed variable data and reconstructed real load data.
3. The method for training a load prediction model for an electric vehicle charging station according to claim 2, wherein a preset model is used to sequentially predict each type of historical variable data at each moment individually to generate multiple sets of predictions Methods of loading data include:

   Each type of reconstructed variable data at each moment is input into the cyclic gate unit network model in turn, and the cyclic gate unit network model outputs multiple sets of predicted load data.
4. The method for training a load prediction model for an electric vehicle charging station according to claim 1, wherein the reinforcement learning method is a Q-learning method, using multiple sets of the predicted load data and the real load data to treat The method of training the weight data set of the trained load forecasting model includes:

   Constructing and initializing a state matrix and an action matrix, wherein the state matrix is composed of the weight data group to be trained, and the action matrix is composed of the variation of the weight;

   The state matrix executes the action matrix according to the preset strategy to update the weight data set;

   Calculate a loss function according to the updated weight data set, multiple sets of predicted load data and the reconstructed real load data, and calculate a reward factor according to the loss function;

   Update the state matrix and the action matrix according to the reward factor;

   Repeat the above steps until the iteration conditions are met.
5. The method for training a load prediction model for an electric vehicle charging station according to claim 4, wherein the calculation formula of the loss function is:

   

   in,

   

   w ₁ , w ₂ ... w _i are weight data groups to be trained, X ₁ , X ₂ ... X _i are multi-class historical variable data, f(X ₁ ), f(X ₂ )...f(X _i ) are multiple Group predicted load data, Y is the real load data, N is the number of groups of historical variable data.
6. A training device for a load prediction model of an electric vehicle charging station, characterized in that the training device comprises:

   an acquisition module, used for acquiring historical load status data of the charging station at several moments, the historical load status data at each moment including multiple types of historical variable data and corresponding real load data;

   The forecasting module is used to make separate forecasts for each type of historical variable data at each moment in turn to generate multiple sets of forecasted load data;

   The training module is used for training the weight data groups of the load prediction model to be trained according to the reinforcement learning method and using multiple sets of the predicted load data and the real load data.
7. The training device for a load prediction model of an electric vehicle charging station according to claim 6, wherein the training device further comprises:

   The data reconstruction module is used for performing phase space reconstruction on the multiple types of the historical load state data to generate multiple types of reconstructed variable data and reconstructed real load data.
8. The training device for a load prediction model of an electric vehicle charging station according to claim 6, wherein the training module comprises:

   an initialization unit, configured to construct and initialize a state matrix and an action matrix, wherein the state matrix is composed of the weight data group to be trained, and the action matrix is composed of the variation of the weight;

   an execution unit, used to make the state matrix execute the action matrix according to the preset strategy to update the weight data group;

   a calculation unit, configured to calculate a loss function according to the updated weight data group, multiple groups of predicted load data and the reconstructed real load data, and calculate a reward factor according to the loss function;

   an update unit, configured to update the state matrix and the action matrix according to the reward factor.
9. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a training program for a load prediction model of an electric vehicle charging station, and the training program for the load prediction model of an electric vehicle charging station is When executed by the processor, the method for training a load prediction model for an electric vehicle charging station according to any one of claims 1 to 5 is implemented.
10. A computer device, characterized in that the computer device comprises a computer-readable storage medium, a processor, and a training program for a load prediction model of an electric vehicle charging station stored in the computer-readable storage medium, the When the training program for the load prediction model for the electric vehicle charging station is executed by the processor, the training method for the load prediction model for the electric vehicle charging station according to any one of claims 1 to 5 is implemented.

Images