CN113902183A - BERT-based non-invasive distribution room charging pile state monitoring and electricity price adjusting method - Google Patents

Abstract

The invention relates to the electric vehicle charging management technology and aims to provide a BERT-based non-invasive distribution room charging pile state monitoring and electricity price adjusting method. The method comprises the following steps: acquiring total power historical data and charging pile state historical data of a transformer substation area as training samples; building a BERT model, which sequentially comprises an embedding layer, a transform layer and an output layer from front to back; determining a loss function for training; training a BERT model by using a gradient descent algorithm; historical total power data of the intelligent electric meters in the transformer substation area are input into the trained BERT model, historical working state data and historical charging power data of each charging pile are obtained, and state monitoring of the charging piles in the transformer area is achieved. The method can be directly applied to the existing charging pile without updating the hardware equipment of the charging pile; the design and production cost of the charging pile are reduced, and the charging pile is more economical and efficient than the traditional method. The charging price can be adjusted in real time to guide a user to change electricity utilization habits, so that the peak clipping and valley filling of a power grid are facilitated, and the electric energy utilization rate is improved.

Description

BERT-based non-invasive distribution room charging pile state monitoring and electricity price adjusting method

Technical Field

The invention relates to an electric vehicle charging management technology, in particular to a non-invasive load detection and charging capacity prediction method based on BERT (bidirectional Encoder recovery from converters), and a method for adjusting charging price of an electric vehicle in real time by using the method.

Technical Field

In recent years, with the rapid development of electric vehicles and related technologies, more and more electric vehicles are connected to a power grid, and the complexity of the load of the power grid is increased. Through the electric vehicle charging capacity prediction algorithm, data support is provided for formulation of the real-time charging electricity price of the electric vehicle, and charging scheduling of the electric vehicle is facilitated. Most of the existing charge capacity prediction methods are realized by electric vehicle charge information acquired by related electrical data detection equipment arranged on a charge pile, so that certain requirements are configured on the hardware of the charge pile, and the existing charge capacity prediction methods are not economical enough.

Disclosure of Invention

The technical problem to be solved by the invention is to overcome the defects in the prior art and provide a non-invasive method for monitoring the charging pile state and adjusting the electricity price of a distribution room based on BERT.

In order to solve the technical problem, the solution of the invention is as follows:

the method for monitoring the state of the charging pile in the non-invasive distribution area based on the BERT comprises the following steps:

(1) acquiring total power historical data and charging pile state historical data of a transformer substation area as training samples;

(2) building a BERT model, which sequentially comprises an embedding layer, a transform layer and an output layer from front to back;

(3) determining a loss function for training;

(4) training the BERT model by using a gradient descent algorithm, which specifically comprises the following steps:

(4.1) randomly initializing model parameters;

(4.2) transmitting the training samples into a BERT model to obtain output;

(4.3) calculating the loss according to the loss function;

(4.4) for each neuron generating an error, adjusting model parameters to reduce the error;

(4.5) repeating steps (4.2) to (4.4) until the loss converges;

(5) historical total power data of the intelligent electric meters in the transformer substation area are input into the trained BERT model, historical working state data and historical charging power data (curves) of the charging piles are obtained, and monitoring of the charging pile state in the transformer area is achieved.

The invention further provides a method for further predicting the residual charging capacity of the transformer substation area by using the BERT-based non-invasive area charging pile state monitoring method, which comprises the following steps:

(1) building and training a fully-connected feedforward neural network FFN;

(2) deploying the trained BERT model and FFN model to an intelligent electric meter, reading total power data of a transformer substation area in real time by the intelligent electric meter, inputting the total power data into the trained BERT model, and acquiring working state data of a charging pile in real time to form historical charging power data (curve);

(3) calculating historical residual capacity data of the transformer area according to the historical total power of the transformer area and the inherent capacity of the transformer area read by the intelligent electric meter; calculating to obtain the historical residual charging capacity of the transformer substation area by combining the historical charging power data obtained in the step (2); and inputting the result into the trained FFN model to obtain the predicted future residual charging capacity.

The invention also provides a method for further adjusting the real-time charging electricity price by utilizing the method for predicting the residual charging capacity of the transformer substation area, which comprises the following steps:

(1) calculating the ratio of the historical/future residual charging capacity of the transformer area to the total transformation capacity of the transformer area within a certain time step according to the historical residual charging capacity and the predicted future residual charging capacity of the transformer area, and recording the ratio as { H }_t-T,H_t-T+1,...,H_t-1,H_tAnd { F }_t,F_t+1,...,F_t+T-1,F_t+T}; wherein H_t＝F_tThe remaining charge capacity is a ratio of the remaining charge capacity at the present moment;

(2) calculate real-time electricity price P that charges that electric automobile used charging pile according to following formula_t：

Wherein rho is an electricity price adjustment constant; k is the number of the time points taken into account, and K time points are considered for history and future residual charging capacity; h_kThe remaining charge capacity ratio of the kth time point before the current time; f_kTo a predicted current timeThe remaining charge capacity at the kth time point after the moment is counted; α ∈ (0,1) is a historical discount factor, γ ∈ (0,1) is a future discount factor, and represents how much charge capacity is left at the previous/next time considered at the current time;

(3) and taking the real-time charging electricity price obtained by calculation as a charging calculation basis for the electric automobile to use the charging pile.

Description of the inventive principles:

under the general condition, all need to install relevant sensor and metering device additional on filling electric pile to the state detection of the electric pile of transformer substation, need more hardware cost.

The non-invasive load detection is a load detection method for obtaining the internal load power data by analyzing the electrical information data of the power load inlet, and has the advantages of economy and cleanness. Nowadays, the energy internet is more and more popular, and the non-invasive load detection is more and more favored by manufacturers and researchers. However, in the related research on non-intrusive load detection, the technical scheme of the related research is to estimate the activities and energy consumption conditions of various other electric devices in the facility by measuring the total electric energy information of the accessed facility, so that the application scene of the technology is basically limited to providing an auxiliary detection means for the research on power consumption control of the electric devices in the facility.

The invention breaks through the inertial thought developed by the technology in the industry, abandons the traditional method of hardware modification of the charging pile, and obtains the electric energy information of the charging pile by using the non-invasive load detection technology, so that the data metering equipment and the sensor are not required to be additionally arranged for monitoring the state of the existing charging pile, and the economical efficiency is improved; meanwhile, the extracted data is used for capacity prediction and power price adjustment, and the instantaneity and the prospect of power price adjustment are improved. The non-intrusive load detection method and the electric vehicle charging capacity prediction are introduced into the real-time adjustment of the charging electricity price of the electric vehicle, and compared with the traditional technical improvement scheme, the method is more prospective and more economic.

The bert (bidirectional Encoder retrieval from transforms) model is usually used for natural language processing tasks, and has a good recognition capability for time series data such as language characters. However, compared with language and character data, the transformer substation area data is more fluctuating and diverse, and the task of extracting the charging pile data from the transformer substation area data cannot be well completed by directly using the BERT model. The invention improves the loss function of the model, so that the model can better fulfill the requirements.

Meanwhile, the traditional electricity price making method generally only considers historical factors (such as historical power generation amount, historical load and the like) and has no prospect. The invention provides a charging price adjustment method considering historical and future residual charging capacity at the same time, and designs a related price adjustment formula, so that the method is more prospective and more reasonable than the traditional method. The future residual charging capacity is obtained by predicting the non-intrusive historical residual charging capacity through the full-connection feedforward neural network FFN.

Based on the improvement, according to total power data of the intelligent electric meters in the transformer substation area, historical working state information of the charging pile of the electric automobile in the transformer substation area is extracted by using a BERT-based non-invasive load detection method; predicting future residual charging capacity by using a fully-connected feed-forward neural network (FFN) according to the historical residual capacity of the transformer substation and the extracted historical charging load information; and finally, formulating a real-time charging electricity price mechanism of the electric automobile according to the historical residual charging capacity and the predicted future residual charging capacity of the charging pile.

Compared with the prior art, the invention has the beneficial effects that:

(1) the non-invasive load detection method is introduced into the electric vehicle charging capacity prediction algorithm, so that the algorithm for obtaining the electric vehicle charging information does not depend on the charging pile to install the electrical data acquisition equipment, can be directly applied to the existing charging pile, and does not need to update the charging pile hardware equipment; the design and production cost of the charging pile are reduced, and the charging pile is more economical and efficient than the traditional method.

(2) After offline training and deployment are carried out, charging power information of the charging pile can be directly obtained online in real time by reading total power data of a transformer substation area, and residual charging capacity information of a predicted future time is updated in real time, so that the method is more real-time and efficient than the traditional algorithm.

(3) The real-time electricity price mechanism of the electric automobile simultaneously considers the historical residual charging capacity and the future residual charging capacity within a certain time step length, and compared with the traditional method, the real-time electricity price mechanism of the electric automobile is more prospective and the established electricity price is more reasonable. The charging price can be adjusted in real time to guide a user to change electricity utilization habits, so that the peak clipping and valley filling of a power grid are facilitated, and the electric energy utilization rate is improved.

Drawings

FIG. 1 is a block diagram of a BERT-based non-intrusive load detection model;

FIG. 2 is a diagram of a fully connected feed forward neural network (FFN) prediction model.

Detailed Description

The invention is further described below with reference to the accompanying drawings and specific embodiments, where an implementation scenario is a substation area where electric vehicle charging piles are built, and the substation area is provided with an intelligent summary table. The invention comprises three phases: firstly, extracting historical working state information of a charging pile of an electric vehicle in a transformer area by using a BERT-based non-invasive load detection method according to total power data of intelligent electric meters in the transformer area; secondly, predicting future residual charging capacity by using a fully-connected feed-forward neural network (FFN) according to the historical residual capacity of the transformer substation and the extracted historical charging load information; and finally, formulating a real-time charging electricity price mechanism of the electric automobile according to the historical residual charging capacity and the predicted future residual charging capacity of the charging pile.

The specific operation of each stage is as follows:

a first part: BERT-based non-invasive load detection method

1. Acquiring total power data of a transformer substation area and charging pile state data as training samples for model input and output;

2. building a BERT model (as shown in figure 1), and sequentially comprising an embedding layer, a transform layer and an output layer from front to back;

(2.1) embedding layer

The embedding layer firstly extracts the characteristics of input data through a convolutional neural network and inputs the characteristics into the hiding layer, reduces the length of an input sequence by half through a square average pooling operation, and finally adds the input sequence with a learnable position embedding matrix (capturing sequence position coding) and outputs the input sequence to the next layer. The calculation formula of the embedding layer is:

Embedding(X)＝LPPooling(Conv(X))+E_pose

where X is the input matrix, LPPooling (. degree.) is the square mean pooling operation, Conv (. degree.) is the convolutional neural network, E_poseRepresenting a learnable position embedding matrix.

(2.2) Transformer layer

Embedding the layer output matrix into the bidirectional Transformer layer, which is composed of multiple layers of transformers and multiple self-attentions in each layer. A single self-entry can be represented by a linear transformation of the input matrices Q (Query), K (Key), and V (value), with the formula:

wherein d is_kDimensions of Q and K, i.e., hidden layer size; softmax (.) is a normalized exponential function.

Integrating multi-head attentions by a plurality of self-attentions, namely, making the process of calculating the self-attentions for a plurality of times, and splicing output matrixes, wherein the formula is as follows:

MultiHead(Q,K,V)＝Concat(head₁,head₂,…,head_h)W^O

wherein Concat (.) is a matrix splicing function; w^OIs a weight matrix; w_i ^Q,W_i ^K,W_i ^VTo linearly map the matrix, the inputs are mapped to different spaces.

(2.3) output layer

The output of the transform layer is first input to a position full connection feedforward network (PF) of the output layerFN) of the formula: pffn (x) ═ GELU (0, XW)₁+b₁)W₂+b₂

Wherein GELU (. eta.) is activation function, W_i,b_iIs a network parameter matrix and X is a network input matrix.

And then obtaining the final output through a multilayer perceptron (MLP), wherein the final output comprises an deconvolution layer and two linear layers, and the formula is as follows:

Out(X)＝Tanh(Deconv(X)W₁+b₁)W₂+b₂

where Tanh (. eta.) is the activation function, Deconv (. eta.) is the deconvolution network, W_i,b_iIs a network parameter matrix and X is a network input matrix.

3. Determining a loss function for training, which is as follows:

wherein,

x∈[0,1]respectively representing a model prediction output value and a normalization value of the real charging pile power;

s_ie { -1,1} is the predicted on-off state and the actual on-off state of the charging pile; t is the total time step; o is the time step length which meets the condition that the actual on-off state of the charging pile is on or the model state is wrong in prediction; both tau and lambda are hyper-parameters of the model, so as to reduce absolute errors; d_KL(.) is a relative entropy function; softmax (.) is a normalized exponential function; log (.) is a log-based function of 10; exp (.) is an exponential function with a constant e as the base.

4. Training a model by using a gradient descent algorithm, and specifically comprising the following steps of:

(4.1) random initialization of model parameters, i.e. weights w_iAnd deviation b_i；

(4.2) transmitting the input data into the model to obtain output;

(4.3) calculating the loss L according to the loss function;

(4.4) for each neuron that produces an error, adjusting the model parameters to reduce the error according to:

(4.5) repeating steps (4.2) to (4.4) until the loss converges.

5. Historical total power data of the intelligent electric meters in the transformer substation area are input into the trained BERT model in real time, historical working state data and historical charging power data (curves) of all charging piles are obtained, and non-invasive load detection is achieved.

A second part: method for predicting residual charging capacity of transformer substation area

1. Building and training a fully connected feed forward neural network (FFN) for predicting future residual charge capacity;

(1) acquiring historical residual capacity of a transformer substation area and historical charging power data of a charging pile, and calculating to obtain the historical residual charging capacity of the transformer substation area; segmenting the data samples to form training data for the FFN model;

(2) an FFN model is built, as shown in fig. 2, which includes a convolution layer and two linear layers, and the calculation formula is:

Out(X)＝Tanh(conv(X)W₁+b₁)W₂+b₂

where Tanh (. eta.) is the activation function, conv (. eta.) is the convolutional network, W_i,b_iX is the network input for the network parameter.

(3) Determining a loss function for training, which is as follows:

wherein,

x_irespectively representing the model output value and the true remaining charge capacity, T being the time step.

(4) Training a model by using a gradient descent algorithm, and specifically comprising the following steps of:

(4.1) random initialization of model parameters, i.e. weights w_iAnd deviation b_i；

(4.2) transmitting the input data into the model to obtain output;

(4.3) calculating the loss L according to the loss function;

(4.4) for each neuron that produces an error, adjusting the model parameters to reduce the error according to:

(4.5) repeating steps (4.2) to (4.4) until the loss converges.

2. And deploying the trained BERT model and FFN model to an intelligent electric meter, reading the total power data of the transformer substation area in real time by the intelligent electric meter, inputting the total power data into the trained BERT model, and acquiring the working state data of the charging pile in real time to form historical charging power data (curve).

3. Calculating historical residual capacity data of the transformer area according to the historical total power of the transformer area and the inherent capacity of the transformer area read by the intelligent electric meter, and calculating the historical residual charging capacity of the transformer area according to the historical charging power data obtained in the step 2; inputting the current data into a trained FFN model to obtain the predicted future residual charge capacity.

And a third part: mechanism for formulating charging real-time electricity price of electric automobile

1. Calculating the ratio of the historical/future residual charging capacity of the transformer area to the total transformation capacity of the transformer area within a certain time step according to the historical residual charging capacity and the predicted future residual charging capacity of the transformer area, and recording the ratio as { H }_t-T,H_t-T+1,...,H_t-1,H_tAnd { F }_t,F_t+1,...,F_t+T-1,F_t+T}; wherein H_t＝F_tThe remaining charge capacity is a ratio of the remaining charge capacity at the present moment;

2. calculate real-time electricity price P that charges that electric automobile used charging pile according to following formula_t：

Wherein rho is an electricity price adjustment constant; h_kIs the ratio of the remaining charge capacity at the kth time point before the current time, F_kThe predicted remaining charge capacity ratio of the kth time point after the current time; k is the number of the time points taken into account, and K time points are considered for history and future residual charging capacity; α ∈ (0,1) is a historical discount factor, γ ∈ (0,1) is a future discount factor, and represents how much charge capacity is left at the previous/next time considered at the current time;

(3) and taking the real-time charging electricity price obtained by calculation as a charging calculation basis for the electric automobile to use the charging pile.

The electricity price mechanism considers the historical residual charging capacity and the future residual charging capacity in a certain time step length simultaneously, and real-time electricity prices are formulated according to the historical residual charging capacity and the future residual charging capacity, so that electricity prices in peak periods and valley periods are increased, electricity prices in valley periods are reduced, users are guided to reduce electricity consumption in peak periods, electricity consumption in valley periods is increased, power grid peak clipping and valley filling are facilitated, and the electric energy utilization rate is improved.

Claims (7)

1. A BERT-based non-invasive method for monitoring the charging pile state in a distribution room is characterized by comprising the following steps:

(1) acquiring total power historical data and charging pile state historical data of a transformer substation area as training samples;

(2) building a BERT model, which sequentially comprises an embedding layer, a transform layer and an output layer from front to back;

(3) determining a loss function for training;

(4) training the BERT model by using a gradient descent algorithm, which specifically comprises the following steps:

(4.1) randomly initializing model parameters;

(4.2) transmitting the training samples into a BERT model to obtain output;

(4.3) calculating the loss according to the loss function;

(4.4) for each neuron generating an error, adjusting model parameters to reduce the error;

(4.5) repeating steps (4.2) to (4.4) until the loss converges;

(5) historical total power data of the intelligent electric meters in the transformer substation area are input into the trained BERT model, historical working state data and historical charging power data of each charging pile are obtained, and state monitoring of the charging piles in the transformer area is achieved.

2. The method according to claim 1, wherein the BERT model specifically comprises:

(2.1) embedding layer

The embedding layer firstly extracts the characteristics of input data through a convolutional neural network and inputs the characteristics into the hiding layer, then the length of an input sequence is halved by adopting square average pooling operation, and finally the input sequence is added with a learnable position embedding matrix and output to the next layer;

(2.2) Transformer layer

Embedding the layer output matrix into a bidirectional Transformer layer, wherein the bidirectional Transformer layer consists of a plurality of layers of transformers and a plurality of self-attentions in each layer; a single self-entry is represented by a linear transformation of the input matrices q (query), k (key), and v (value); integrating multi-head attentions by a plurality of self-attentions, namely, performing a self-attention calculation process for a plurality of times, splicing output matrixes, and mapping input to different spaces;

(2.3) output layer

The output of the Transformer layer is firstly input to a position full-connection feedforward network PFFN of an output layer; and then the final output is obtained through a multi-layer perceptron MLP comprising an deconvolution layer and two linear layers.

3. The method according to claim 1, wherein the training loss function in step (3) is specifically:

wherein,

respectively representing a model prediction output value and a normalization value of the real charging pile power;

the predicted switch state and the actual switch state of the charging pile are obtained; t is the total time step; o is the time step length which meets the condition that the actual on-off state of the charging pile is on or the model state is wrong in prediction; both tau and lambda are hyper-parameters of the model, so as to reduce absolute errors; d_KL(.) is a relative entropy function; softmax (.) is a normalized exponential function; log (.) is a log-based function of 10; exp (.) is an exponential function with a constant e as the base.

4. The method for further predicting the remaining charging capacity of a substation area by using the BERT-based non-intrusive area charging pile state monitoring method of claim 1, characterized by comprising the following steps:

(1) building and training a fully-connected feedforward neural network FFN;

(2) deploying the trained BERT model and FFN model to an intelligent electric meter, reading total power data of a transformer substation area in real time by the intelligent electric meter, inputting the total power data into the trained BERT model, and acquiring working state data of a charging pile in real time to form historical charging power data;

(3) calculating historical residual capacity data of the transformer area according to the historical total power of the transformer area and the inherent capacity of the transformer area read by the intelligent electric meter; calculating to obtain the historical residual charging capacity of the transformer substation area by combining the historical charging power data obtained in the step (2); and inputting the result into the trained FFN model to obtain the predicted future residual charging capacity.

5. The method according to claim 4, characterized in that said step (1) comprises in particular:

(1.1) acquiring historical residual capacity of a transformer substation area and historical charging power data of a charging pile, and calculating to obtain the historical residual charging capacity of the transformer substation area; segmenting the data sample to form training data;

(1.2) building an FFN model, which comprises a convolution layer and two linear layers;

(1.3) determining a loss function for training;

(1.4) training the FFN model by using a gradient descent algorithm, which specifically comprises the following steps:

(1.4.1) randomly initializing model parameters;

(1.4.2) transmitting the input data into the model to obtain output;

(1.4.3) calculating a loss according to the loss function;

(1.4.4) for each neuron that produces an error, adjusting model parameters to reduce the error:

(1.4.5) repeating steps (1.4.2) to (1.4.4) until the loss converges.

6. The method according to claim 5, wherein the training loss function in step (1.3) is specifically:

wherein,

x_irespectively representing the model output value and the true remaining charge capacity, T being the time step.

7. The method for predicting the remaining charging capacity of the substation area according to claim 4 is further used for adjusting the real-time charging electricity price, and is characterized by comprising the following steps:

(1) historical remaining charge capacity and predicted future remaining based on substation areaThe residual charging capacity is calculated to obtain the ratio of the station area history/future residual charging capacity of a fixed time step length in a period of time to the total transformation capacity of the station area, and is recorded as { H_t-T,H_t-T+1,...,H_t-1,H_tAnd { F }_t,F_t+1,...,F_t+T-1,F_t+T}; wherein H_t＝F_tThe remaining charge capacity is a ratio of the remaining charge capacity at the present moment;

(2) calculate real-time electricity price P that charges that electric automobile used charging pile according to following formula_t：

Wherein rho is an electricity price adjustment constant; k is the number of the time points taken into account, and K time points are considered for history and future residual charging capacity; h_kThe remaining charge capacity ratio of the kth time point before the current time; f_kThe predicted remaining charge capacity ratio at the kth time point after the current time is obtained; α ∈ (0,1) is a historical discount factor, γ ∈ (0,1) is a future discount factor, and represents how much charge capacity is left at the previous/next time considered at the current time;

(3) and taking the real-time charging electricity price obtained by calculation as a charging calculation basis for the electric automobile to use the charging pile.

Images