A deep learning multi-model electric vehicle charging station short-term load prediction method

Abstract

The application discloses a kind of deep learning multi-model electric vehicle charging station short-term load prediction method, including CNN main module, LSTM main module and Transformer main module, fusion CNN obtains the load curve feature, the context information obtained by LSTM and the global information obtained by multi-attention mechanism in Transformer;The method disclosed in the application comprehensively the results of two input-output modes, one is input as multi-dimensional historical data at the same time, output as prediction data at the same time, one is input as adjacent multi-dimensional historical and prediction data, output as prediction data at adjacent time, improve the accuracy of electric vehicle charging station short-term load prediction.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and in particular to a deep learning multi-model method for short-term load forecasting of electric vehicle charging stations. Background Technology

[0002] When electric vehicle charging stations are operating, the harmonics they generate can affect the voltage and current quality of the power grid, thus impacting the normal operation of other electrical equipment. It is necessary to monitor the charging patterns of electric vehicle charging stations in a timely manner to balance the three-phase charging load and reduce the impact of harmonics. Furthermore, the varying charging demands and charging time distribution of different numbers of electric vehicles will create different loads on the power system. Understanding the patterns of electric vehicle charging loads is beneficial for guiding off-peak charging of electric vehicles and for enabling the power grid to expand its capacity in a timely manner to meet user electricity demand. Load forecasting is one of the effective technologies for understanding user electricity consumption in advance.

[0003] Electricity load forecasting includes residential load forecasting, industrial load forecasting, and regional load forecasting. Early load forecasting methods were mainly based on traditional machine learning, such as gray-scale forecasting, SVM forecasting, and neural network forecasting. Relatively speaking, after introducing load influencing factors, the accuracy of some regular load forecasts was better. For example, residential load is closely related to residents' living habits, weather, and holidays, and the forecast results are highly accurate. However, electric vehicle charging load forecasting is relatively difficult, mainly due to the mobility of electric vehicles, the rapid growth in the number of electric vehicles, and the high randomness of charging time. Traditional machine learning methods perform poorly in forecasting electric vehicle charging station load. With the application of deep learning technology, electric vehicle charging load forecasting methods have shifted from traditional machine learning methods to deep learning methods, mainly including CNN-based forecasting, RNN-based forecasting, LSTM-based forecasting, and Transformer-based forecasting. Summary of the Invention

[0004] To overcome the shortcomings of the existing technology, the purpose of this invention is to provide a deep learning multi-model method for short-term load forecasting of electric vehicle charging stations. The load forecasting method proposed in this invention includes a CNN main module, an LSTM main module, and a Transformer main module, which integrates the load curve features obtained by CNN, the contextual information obtained by LSTM, and the global information obtained by the multi-attention mechanism in Transformer. The proposed method combines the results of two input-output methods: one is inputting multidimensional historical data at the same moment and outputting predicted data at the same moment, and the other is inputting adjacent multidimensional historical and predicted data and outputting predicted data at adjacent moments, so as to make the short-term load forecasting of electric vehicle charging stations more accurate.

[0005] The technical solution of this invention to solve its technical problem is:

[0006] A deep learning multi-model method for short-term load forecasting of electric vehicle charging stations includes:

[0007] A convolutional neural network (CNN) main module is established, consisting of a single convolutional layer. Input data is fed into the CNN main module for preliminary feature extraction, yielding preliminary features X1[x1,x2,…,x]. n ];

[0008] Establish a Long Short-Term Memory (LSTM) network main module, consisting of n LSTM units, and process the initial features X1[x1,x2,…,x...]. n The input is cascaded into LSTM units to capture the temporal relationships between features, resulting in time series prediction features F2[f1,f2,…,f n ];

[0009] The main Transformer module is established, which mainly consists of a Transformer encoder. It converts the time series prediction features F2[f1,f2,…,f…] into… n The Transformer main module is introduced to extract context information from time series data and perform transformation and encoding to obtain intermediate features F3;

[0010] A fully connected layer is established, and the intermediate feature F3 is input into a fully connected layer to obtain the prediction result.

[0011] Furthermore, the network model of the method is trained with two input and output methods to obtain two parameters P1 and P2 of the same model;

[0012] The first input-output method: The input is the data at time n over the past 96 days, and the output is the data at time n. Under this input-output method, the model is trained to obtain the model parameters P1.

[0013] The second input-output method: the input is the data from the past 96 consecutive time steps, and the output is the data from the current time step n. Under this input-output method, the model is trained to obtain the model parameters P2.

[0014] Furthermore, the model parameters P1 and P2 are further processed together. To obtain the load forecast value X for a certain day, the historical data required by model parameter P1 is input, and the forecast value obtained by model parameter P1 is X'. The historical data required by model parameter P2 is input, and the forecast value obtained by model parameter P2 is X”. The two are weighted and averaged, X = αX' + (1-α)X”, (0≤α≤1), to obtain the load forecast value X for the desired day.

[0015] The beneficial effects of this invention are:

[0016] The load forecasting method proposed in this invention includes a CNN main module, an LSTM main module, and a Transformer main module. It integrates the load curve features obtained by CNN, the contextual information obtained by LSTM, and the global information obtained by the multi-attention mechanism in Transformer. The proposed method combines the results of two input-output methods: one is to input multidimensional historical data at the same moment and output the predicted data at the same moment, and the other is to input adjacent multidimensional historical and predicted data and output the predicted data at adjacent moments. This can effectively improve the accuracy of short-term load forecasting for electric vehicle charging stations. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the short-term load forecasting method for electric vehicle charging stations according to the present invention.

[0018] Figure 2 These are schematic diagrams illustrating two input / output methods of the present invention;

[0019] Figure 3 Schematic diagram of an LSTM unit;

[0020] Figure 4 This is a schematic diagram of one layer of the Transformer encoder. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings.

[0022] Taking the short-term load forecasting of an electric vehicle charging station as an example, assume that the charging station's charging load is collected every 15 minutes to form historical data, which is then used as sample data for learning. Figure 1 As shown, a deep learning multi-model method for short-term load forecasting of electric vehicle charging stations consists of three main modules: a convolutional neural network (CNN), a long short-term memory network (LSTM), and a Transformer. The CNN module consists of one convolutional layer, the LSTM module consists of n LSTM units, and the Transformer module consists of a Transformer encoder. For ease of description, the historical data for k days is recorded as [y]. 1,1 ,y 1,2 ,y 1,3 ,……,y 1,94 ,y 1,95 ,y 1,96 ],[y 2,1 ,y 2,2 ,y 2,3 ,……,y 2,94 ,y 2,95 ,y 2,96 ],[y 3,1 ,y 3,2 ,y3,3 ,……,y 3,94 ,y 3,95 ,y 3,96 ],......,[y k,1 ,y k,2 ,y k,3 ,……,y k,94 ,y k,95 ,y k,96 Let [y] be the data for predicting 1 day. k+1,1 ,y k+1,2 ,y k+1,3 ,……,y k+1,94 ,y k+1,95 ,y k+1,96 The process includes the following steps:

[0023] Step S1: Establish the main module of the Convolutional Neural Network (CNN), which consists of one convolutional layer, and input data D... Input [y i1,j1, y i2,j2 ,y i3,j3 ,……,y i94,j94 ,y i95,j95 ,y i96,j96 The input is fed into the main CNN module for preliminary feature extraction, resulting in preliminary features X1[x1,x2,…,x]. n The details are as follows:

[0024] X1 = H 3×3 (D Input (1)

[0025] Among them, H 3×3 (·) indicates a 3×3 convolution operation.

[0026] Step S2: Establish the main module of the Long Short-Term Memory (LSTM) network, which consists of n LSTM units. The initial features X1[x1,x2,…,x] are then processed. n The input is cascaded into LSTM units to capture the temporal relationships between features, resulting in time series prediction features F2[f1,f2,…,f n The details are as follows:

[0027] f i =L i (x i (2) i = 1, 2, ..., n

[0028] Among them, L i (·) The i-th LSTM unit performs time series prediction operation.

[0029] like Figure 3As shown, an LSTM unit consists of three gates: the forget gate, the input gate, and the output gate. Ci represents the unit's long-term memory, and hi represents the unit's short-term memory.

[0030] The forget gate first takes the input x at the current time step t. t h at the previous time step t-1 t-1 By splicing, we get [h] t-1 ,x t Then, a fully connected layer is used for transformation, and finally, the sigmoid function is used for activation to obtain the forget gate value f. t :

[0031] f t =σ(W f ·[h t-1 ,x t ]+b f (3)

[0032] Where t is the current time, σ represents the sigmoid function, and W f For the weight, h t-1 Let b be the hidden state at time t-1. f This is a deviation.

[0033] The input gate first generates the input gate parameters i t Generate a candidate value for updating.

[0034] i t =σ(W i ·[h t-1 x t ]+b i (4)

[0035]

[0036] Update C t-1 The obtained forget gate value f t C obtained from the previous time step t-1 t-1 Multiply, and add the input gate parameter i t Candidate values ​​obtained at the current time step The result of multiplication is the updated C. t As part of the input for the next time step.

[0037]

[0038] The output gate first calculates the output gate parameter o. t For the updated C t Perform tanh activation, multiply the two to get h. tAs part of the input for the next time step t+1, the entire output gate process is designed to generate the unit's short-term memory h. t .

[0039] o t =σ(W o ·[h t-1 x t ]+b o (7)

[0040] h t =o t *tanh(C t (8)

[0041] Step S3: Establish the Transformer main module, which mainly consists of a Transformer encoder, and converts the time series prediction features F2[f1, f2, ..., f... n The Transformer main module is introduced to extract context information from time series data and perform transformation and encoding to obtain intermediate features F3, as follows:

[0042] F3 = H Trans (F2) (9) Wherein, H Trans (·) indicates Transformer encoder operation.

[0043] The Transformer encoder consists of N=6 identical layers, each layer (e.g., ...) Figure 4 (As shown) There are two sub-layers. The first sub-layer is a multi-head self-attention mechanism function, Attention, and the second sub-layer is a fully connected feedforward neural network. Each layer in both sub-layers is surrounded by a residual connection, and each layer contains layer normalization. That is, the output of each sub-layer is LayerNorm(x + Sublayer(x)), where Sublayer(x) is a function implemented by the sub-layer itself. To facilitate these residual connections, the output dimension of all sub-layers and embedding layers in the model is d. model =128.

[0044] Multi-head self-attention can be described as mapping a query (Q) and a set of key-value pairs (K, V) to an output, where the query (Q), keys (K), values ​​(V), and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is calculated by a compatibility function of the query (Q) with the corresponding key (K). The input consists of dimension d. k The query (Q) and key (K) are composed of the values ​​of dimension dk. Calculate the dot product of all queries (Q) and keys (K), and divide each by dk. Finally, the weights of the values ​​(v) are obtained using the softmax function. The specific formula is as follows:

[0045]

[0046] The fully connected feedforward neural network consists of a linear layer, a LayerNorm layer, and a dropout layer in sequence.

[0047] p i =G(L(z) i i = 1, 2, ..., n (11)

[0048] O i =D(L(D(p) i )))i=1,2,…,n (12)

[0049] Where G(·) is the GELU activation function, L(·) is the Linear fully connected layer, D(·) is the Dropout function, and z i It is the input of the feedforward network, O i It is the output of the feedforward network.

[0050] Step S4: Establish a fully connected layer by inputting the intermediate feature F3 into a fully connected layer to obtain the prediction result I. output [x o The details are as follows:

[0051] I output =H FC (F3) (4)

[0052] Among them, H FC (·) indicates a full connection operation.

[0053] Step s5: The network model of the method is trained using two input / output methods to obtain two parameters P1 and P2 of the same model. This treats the method as a deep learning network model, and the network model is trained as follows: Figure 2 As shown, the details are as follows:

[0054] Step S51, First input / output method:

[0055] Method input D Input [X i1，j1 X i2，j2 X i3，j3 ..., X i94，j94 X i95，j95 X i96，j96In the input-output model, j1=j2=j3=......=j94=j95=j96=n, where n=1,2,3,...,94,95,96, which represents the data at the same time i1, i2, i3,...i94,i95,i96 days ago at time n. The output is the data at the current time n. Under this input-output method, the model is trained to obtain the model parameters P1.

[0056] Step S52, Second input / output method:

[0057] Method input D Input [x i1,j1 ,x i2,j2 ,x i3,j3 ,……,x i94,j94 ,x i95,j95 ,x i96,j96 The input consists of data from the past 96 consecutive time points, and the output is the data from the current time point n. With this input-output method, the model is trained to obtain the model parameters P2.

[0058] Step S6: Combine the two outputs, as detailed below:

[0059] When predicting the load on a given day, historical data from the past 96 days is needed. Under model parameter P1, the data at time 96 is predicted, X' = [x' k+1,1 ,x' k+1,2 ,x' k+1,3 ,……,x' k+1,94 ,x' k+1,95 ,x' k+1,96 Under model parameters P2, using the data y1, y2, y3, ..., y94, y95, y96 from the past 96 time points, we obtain x” k+1,1 Then use x” k+1,1 Update the data for the past 96 time points: y2, y3, ..., y94, y95, y96, x” k+1,1 , to obtain x” k+1,2 Then use x” k+1,1 ,x” k+1,2 Update the data for the past 96 time points: y3,......,y94,y95,y96,x” k+1,1 ,x” k+1,2 , to obtain x” k+1,3 And so on, we get X” = [x” k+1,1 ,x” k+1,2 ,x” k+1,3 ,……,x” k+1,94 ,x” k+1,95 ,x” k+1,96 ].

[0060] The weighted average of X' and X” is used to obtain X, X=αX'+(1-α)X”, (0≤α≤1), which is the load forecast value for a certain day.

[0061] The above description merely illustrates preferred embodiments of the present invention, and while the description is relatively specific and detailed, it should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications, improvements, and substitutions without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A deep learning multi-model method for short-term load forecasting of electric vehicle charging stations, characterized in that, include: A convolutional neural network (CNN) main module is established, consisting of a single convolutional layer. Input data is fed into the CNN main module for preliminary feature extraction, yielding preliminary features. ; Establish a Long Short-Term Memory (LSTM) network main module, consisting of n LSTM units, to process initial features. By inputting cascaded LSTM units, the temporal relationships between features are captured to obtain time series prediction features. ; The main Transformer module is established, primarily consisting of a Transformer encoder, which transforms time series prediction features. The Transformer main module is introduced to extract contextual information from time-series data and perform transformations and encoding to obtain intermediate features. ; Establish a fully connected layer to integrate intermediate features Input a fully connected layer to obtain the prediction result; In this method, the network model is trained using two input / output methods to obtain two parameters P1 and P2 for the same model. The first input-output method: The input is the data at time n over the past 96 days, and the output is the data at time n. Under this input-output method, the model is trained to obtain the model parameters P1. The second input-output method: The input is the data from the past 96 consecutive time steps, and the output is the data from the current n time steps. Under this input-output method, the model is trained to obtain the model parameters P2. Further processing of model parameters P1 and P2 is performed to obtain the load forecast value X for a certain day. The historical data required by model parameter P1 is input, and the forecast value obtained by model parameter P1 is X'. The historical data required by model parameter P2 is input, and the forecast value obtained by model parameter P2 is X''. The two are weighted and averaged, X=αX'+(1-α)X'',0≤α≤1, to obtain the load forecast value X for the desired day.

Images