Electric vehicle charging load prediction method and system based on MCBDL-TCN model

Abstract

The invention discloses an electric vehicle charging load prediction method and system based on an MCBDL-TCN model, and is applied to the technical field of charging demand prediction. The method comprises the following steps: obtaining charging demand data of an electric vehicle charging station and corresponding meteorological data, converting the data into time sequence data, constructing a charging load prediction model based on a multi-cascade Bayesian deep learning model and a time convolution model, optimizing the weight W and deviation b of LSTM-BDL in combination with the Bayesian theory, and obtaining the charging load prediction model of the electric vehicle charging station. And training the charging load prediction model, testing the charging load prediction model to determine the optimal layer number of the model, and performing prediction based on the charging load prediction model. The charging load prediction model has excellent performance in the aspects of predicting and solving uncertainty, so that the accuracy and the flexibility of complex time sequence data are improved, contributions can be made to the field of electric vehicle charging load prediction, and a reliable tool is provided for coping with ever-increasing electric vehicle infrastructure requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of charging load prediction, and more specifically to a method and system for predicting charging load of an electric vehicle based on a MCBDL-TCN model. Background Art

[0002] With the deep integration of electric vehicles and distribution networks, the randomness and uncertainty of their charging demand bring new challenges to power dispatching. Therefore, conducting research on charging load forecasting will not only help promote electric vehicle technology, but also ensure the safe and stable operation of the power grid system. In modern complex distribution systems, multi-dimensional factors such as meteorological conditions, holiday arrangements, real-time electricity prices, urban development levels and user habits increasingly affect electricity demand. Machine learning has demonstrated excellent performance in the field of recent load forecasting with its adaptive modeling capabilities. Although it has obvious advantages over traditional methods, current intelligent algorithms still need to be improved in dealing with forecasting uncertainties and grasping long-term changing trends. The development of electric vehicle charging load forecasting technology and its application system based on the MCBDL-TCN model has become a key issue that needs to be broken through in this field. Summary of the invention

[0003] Based on this, the present invention provides an electric vehicle charging demand prediction technology and its application platform integrating the MCBDL-TCN architecture, which significantly improves the prediction accuracy by integrating multi-level Bayesian learning and time series convolution algorithm.

[0004] In order to achieve the above object, the present invention provides the following technical solutions: S1. Obtain charging demand data of electric vehicle charging stations and corresponding meteorological data; S2, converting the charging demand data and meteorological data of electric vehicle charging stations into time series data, and dividing them into training set, validation set and test set; S3. Construct a charging load prediction model based on a multi-cascade Bayesian deep learning model and a time convolution model; The MCBDL architecture realizes the progressive extraction of data features by connecting multiple LSTM-BDL units in series. First, LSTM-BDL is used to process the time series input to obtain uncertainty features, and then the uncertainty features are passed to the TCN module for in-depth analysis. This combination design gives full play to the advantages of the two networks, enhances the processing ability of complex time series data, and improves the capture effect of global dependencies and multi-scale features. Finally, the output information of MCBDL and TCN is integrated through the fully connected layer to generate the prediction results. The core component of TCN is RDCCU, which contains two rounds of causal convolution components with the same expansion, Weight Norm layer, activation function, Dropout layer and input residual connection. A multi-layer TCN network architecture is constructed by connecting multiple RDCCUs in series. S4. Optimize the weight W and bias b of LSTM-BDL in combination with Bayesian theory, use standard normal distribution as the prior distribution of weight W and bias b, determine the posterior distribution of weight W and bias b according to Bayesian theory, and process the posterior distribution through variational inference method to obtain the optimal distribution of weight W and bias; S5. Training a charging load prediction model based on the training set and the validation set; S6. Using the test data set, the MAE and RMSE indicators are calculated for MCBDL models of different depths, and the optimal number of layers of the LSTM-BDL model is determined based on the minimum value of MAE and RMSE; S7. Collect historical operation data and environmental parameters of the target charging station, convert this information into a standard time series format, and input it into the trained charging load prediction model to obtain the final load prediction value.

[0005] Optionally, S1 collects the operating parameters of the charging equipment at each station including actual power, transaction power and charging period, and records the power value every 15 minutes; synchronously obtains corresponding meteorological data including precipitation, temperature, air humidity, visibility and wind speed, and performs spline interpolation processing on the meteorological data.

[0006] Optionally, S3 is as follows: After the feature information extracted by MCBDL is passed into TCN, each layer of LSTM-BDL modules continuously analyzes the uncertainty in the data. This design effectively reduces the error accumulation in complex data processing and improves the system's modeling accuracy and robustness for time series variables. The cascade function of the charging load prediction model is expressed as follows: ; In the formula, is the output of LSTM-BDL, l is the number of cycles of LSTM-BDL, n is the multiplicity of the cascade, for l- 1 layer at the current time step t The hidden state output of for l The layer at the current time step t The hidden state output is -1, For l Layer at the previous time step t -1 memory cell status; The original intention of the design of fusing LSTM-BDL and TCN is to integrate their respective advantages: the former's expertise in sequence modeling and uncertainty processing, and the latter's high efficiency in temporal dependency extraction. The MCBDL-TCN architecture constructed in this way enhances the ability to analyze complex data, especially when processing the dynamic characteristics of load data, ensuring the reliability of the prediction results.

[0007] Optionally, the calculation function of LSTM-BDL is: ; ; In the formula, is the current input, σ represents the sigmoid function, and Represent the weight and bias of LSTM-BDL respectively, f t is the forgotten value, o t is the output value, c t For the memory unit, is a candidate memory unit, i t To update the value, h t-1 is the hidden state at the previous moment, h t is the hidden state at the current moment.

[0008] Optionally, the temporal convolution model is composed of multiple residual dilated causal convolution units connected in series, and its unit structure is: ; In the formula, o Represents the result after activation function processing, using Artivation Activation function, x represents the output from the previous layer, F(x ) represents the residual function.

[0009] Optionally, the core of the Residual Dilated Causal Convolution Unit (RDCCU) is the dilated causal convolution, and the mathematical model is expressed as: ; In the formula, x and y Represent the input and output of the convolution respectively, s is the current time step, f The size is K The convolution kernel, iis the convolution operation index, m is the expansion factor, represents time series data, x s is the time step s The input signal value at ; Causal convolution uses only the input data of the current and historical moments for calculation, which not only meets the needs of power load forecasting, but also prevents future information leakage. m , which can expand the receptive field while keeping the depth unchanged and improve the training effect; at the same time, the front layer input is periodically sampled. m The calculation is: ; In the formula, L The computed output length for the convolution operation.

[0010] Optionally, S6 specifically includes: inputting the test set into charging load prediction models with different numbers of layers, and evaluating the prediction performance of each model by calculating MAE and RMSE, and the calculation formula is as follows: ; In the formula, N is the number of tests, i ∈[1, N ], y i For the i The predicted load of the test, y fi For the i The actual load of the test; The optimal number of layers of the LSTM-BDL model is determined based on the minimum values ​​of MAE and RMSE.

[0011] An electric vehicle charging load prediction system based on the MCBDL-TCN model includes the following five modules: a data acquisition module is responsible for collecting charging station demand data and local meteorological data; a data processing module processes the charging demand data; a charging load prediction module is built based on multiple cascaded Bayesian deep learning and time convolution models; a model evaluation module evaluates the prediction performance; and a model optimization module performs model optimization.

[0012] Optionally, the charging load prediction model consists of three parts: a multi-cascade Bayesian deep learning model formed by cascading multiple LSTM-BDLs, a temporal convolution model, and a Linear layer.

[0013] Optionally, the temporal convolution model is composed of multiple residual dilated causal convolution units connected in series, each residual dilated causal convolution unit contains two paths: the main path is composed of two groups of identical structures connected in series, each group of structures is a dilated causal convolution factor, a Weight Norm layer, a ReLu activation function and a Dropout layer; the residual path is a 1×1 convolution layer, and finally the output of the second Dropout layer of the main path is connected to the output of the 1×1 convolution layer of the residual path.

[0014] It can be seen from the above technical solution that compared with the prior art, the present invention provides an electric vehicle charging load forecasting method and system based on the MCBDL-TCN model, which has the following beneficial effects: the present invention combines multi-stage Bayesian deep learning with a temporal convolutional network to capture short-term changes and long-term patterns in electric vehicle charging demand, and uses TCN to improve the model's ability to process long sequences and capture global time-related features. LSTM-BDL is used to extract detailed features from complex time series data and process uncertainties in predictions, with stronger prediction accuracy. It is particularly suitable for applications that require accurate load forecasting, such as power system management and urban planning, and can contribute to the field of electric vehicle charging load forecasting and provide a reliable tool for meeting the growing demand for electric vehicle infrastructure. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0016] Figure 1 This is a flow chart of the electric vehicle charging load prediction method of the present invention; Figure 2 It is a schematic diagram of the MCBDL-TCN model structure of the present invention; Figure 3 It is a schematic diagram of the LSTM-BDL model structure of the present invention; Figure 4 It is a schematic diagram of the structure of the temporal convolution model of the present invention; Figure 5 Schematic diagram of the dilated causal convolution structure of the present invention; Figure 6 Schematic diagram of prediction result evaluation index under different LSTM-BDL layers in an embodiment of the present invention; Figure 7 It is a schematic diagram of load prediction results of the MCBDL-TCN model in an embodiment of the present invention; Figure 8A schematic diagram of load prediction results of the ANN model in an embodiment of the present invention; Fig. 9 This is a schematic diagram of the load prediction results of the LSTM model in an embodiment of the present invention; Fig.10 Schematic diagram of the root mean square error of each model in the embodiment of the present invention; Fig.11 Schematic diagram of mean absolute error of each model in an embodiment of the present invention; Fig.12 Schematic diagram of the mean absolute percentage error of each model in an embodiment of the present invention; Fig.13 Schematic diagram of average loss of QR and MCBDL-TCN on working days and non-working days in an embodiment of the present invention Fig.14 Schematic diagram of overall probability evaluation indicators of QR and MCBDL-TCN in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] The embodiment of the present invention discloses a method for predicting electric vehicle charging load based on the MCBDL-TCN model. Figure 1 As shown, the following steps are included: S1. Obtain charging demand data of electric vehicle charging stations and corresponding meteorological data; S2, converting the charging demand data and meteorological data of electric vehicle charging stations into time series data, and dividing them into training set, validation set and test set; S3. Construct a charging load prediction model based on a multi-cascade Bayesian deep learning model and a time convolution model; like Figure 2As shown in the figure, the MCBDL architecture realizes the progressive extraction of data features by connecting multiple LSTM-BDL units in series. First, LSTM-BDL is used to process the time series input to obtain uncertainty features, and then the uncertainty features are passed to the TCN module for in-depth analysis. This combination design gives full play to the advantages of the two networks, enhances the processing ability of complex time series data, and improves the capture effect of global dependencies and multi-scale features. Finally, the output information of MCBDL and TCN is integrated through the fully connected layer to generate the prediction result. The core component of TCN is RDCCU, which contains two rounds of causal convolution components with the same expansion, Weight Norm layer, activation function, Dropout layer and input residual connection. A multi-layer TCN network architecture is constructed by connecting multiple RDCCUs in series. S4. Optimize the weight W and bias b of LSTM-BDL in combination with Bayesian theory, use standard normal distribution as the prior distribution of weight W and bias b, determine the posterior distribution of weight W and bias b according to Bayesian theory, and process the posterior distribution through variational inference method to obtain the optimal distribution of weight W and bias; S5. Training a charging load prediction model based on the training set and the validation set; S6. Using the test data set, the MAE and RMSE indicators are calculated for MCBDL models of different depths, and the optimal number of layers of the LSTM-BDL model is determined based on the minimum value of MAE and RMSE; S7. Collect historical operation data and environmental parameters of the target charging station, convert this information into a standard time series format, and input it into the trained charging load prediction model to obtain the final load prediction value.

[0019] Furthermore, S1 collects the operating parameters of the charging equipment at each station, including actual power, transaction power and charging period, and records the power value every 15 minutes; synchronously obtains the corresponding meteorological data including precipitation, temperature, air humidity, visibility and wind speed, and performs spline interpolation processing on the meteorological data.

[0020] Furthermore, S3 is as follows: After the feature information extracted by MCBDL is passed into TCN, each layer of LSTM-BDL modules continuously analyzes the uncertainty in the data. This design effectively reduces the error accumulation in complex data processing, while improving the system's modeling accuracy and robustness for time series variables. The cascade function of the charging load prediction model is expressed as follows: ; In the formula, is the output of LSTM-BDL, l is the number of cycles of LSTM-BDL, n is the multiplicity of the cascade, for l-1 layer at the current time step t The hidden state output of for l The layer at the current time step t The hidden state output is -1, For l Layer at the previous time step t -1 memory cell status; The original intention of the design of fusing LSTM-BDL and TCN is to integrate their respective advantages: the former's expertise in sequence modeling and uncertainty processing, and the latter's high efficiency in temporal dependency extraction. The MCBDL-TCN architecture constructed in this way enhances the ability to analyze complex data, especially when processing the dynamic characteristics of load data, ensuring the reliability of the prediction results.

[0021] In the field of time series data analysis, LSTM-BDL and TCN each have their own strengths. With its Bayesian mechanism, LSTM-BDL is good at handling the continuous impact of the sequence and quantifying the uncertainty of the prediction, which makes it robust in the face of data fluctuations. Its gating mechanism can dynamically filter sequence information to ensure accurate grasp of data changes. TCN has unique advantages in long-range information extraction through its extension and causal convolution structure, and effectively obtains global time series associations. Its residual connection design further enhances the training effect of the system.

[0022] The original intention of the design of fusion LSTM-BDL and TCN is to integrate their respective advantages: the former's expertise in sequence modeling and uncertainty processing, and the latter's efficiency in extracting time series dependencies. The MCBDL-TCN architecture constructed in this way enhances the ability to analyze complex data, especially when processing the dynamic characteristics of load data, ensuring the reliability of prediction results.

[0023] Furthermore, the calculation function of LSTM-BDL is: ; ; In the formula, is the current input, σ represents the sigmoid function, and Represent the weight and bias of LSTM-BDL respectively, f t is the forgotten value, o t is the output value, c t For the memory unit, is a candidate memory unit, i t To update the value, ht-1 is the hidden state at the previous moment, h t is the hidden state at the current moment; LSTM-BDL combines the characteristics of long short-term memory network and Bayesian deep learning. Its structure is as follows: Figure 3 As shown in the figure. As a specialized version of RNN, LSTM has shown excellent performance in areas such as load forecasting. Its core components include internal hidden states, and manage information flows through independent memory units and three types of gating mechanisms: forget gates, input gates, and output gates, which are responsible for removing redundant information, introducing new data elements, and controlling the degree of exposure of internal states. This three-gate structure enables memory units to flexibly update and filter information and effectively handle long-term associations in sequences. Based on the serial characteristics of LSTM, a deep network architecture can be constructed by stacking multiple layers. The innovation of LSTM-BDL is to replace the deterministic parameters (weights and biases) of traditional LSTM with a group of random variables and introduce the Bayesian theory framework.

[0024] In the embodiment of the present invention, S5 is specifically: the weight W is represented by a prior distribution such as N(0, 1). Since the network parameters of BNN (including weights and biases) lack clear physical meanings, their prior beliefs are difficult to accurately predict. For such parameter uncertainty problems, using a standard normal distribution is a practical solution, that is, let the prior distribution p(W)∼N(0, 1).

[0025] After determining the prior distribution, the model p ( Y train )| h w ( X train ))'s likelihood function uses the normal distribution N ( h w ( X train ), σ 2 ) describes the noise level σ Keep stable, the relationship between the target LSTM and the model parameter W is expressed as h w ( X train ) indicates that according to Bayes’ theorem, the posterior distribution of the parameter can be expressed as: ; based on p ( W )| h w ( X train )), for the new input point The output value can be predicted. As a random variable, its expression is: ; The embodiment of the present invention uses the variational inference (VI) method to deal with the learning and reasoning problems in BDL. The advantage of VI is that it converts the difficult-to-handle posterior calculation into an approximate distribution form, which is particularly important when dealing with complex network structures and multidimensional data. The core idea is to approximate the posterior distribution through the variational distribution. The specific implementation process includes selecting an approximate variational distribution family controlled by the parameter τ and adjusting τ Optimize the distribution performance, and then minimize the KL divergence between the variational distribution and the target distribution to obtain the optimal distribution of weights W and bias b. The definition of KL divergence is: ; in, q τ ( W ) is an approximate distribution used in variational inference to estimate the posterior distribution of complex models p ( W )| h w ( X train )).

[0026] Furthermore, the temporal convolution model is composed of multiple residual dilated causal convolution units connected in series, and its unit structure is: ; In the formula, o Represents the result after activation function processing, using Artivation Activation function, x represents the output from the previous layer, F(x ) represents the residual function.

[0027] The core component of TCN is the residual dilated causal convolution unit, which consists of two rounds of identically dilated causal convolution components, a Weight Norm layer, an activation function, a Dropout layer, and an input residual connection. A multi-layer TCN network architecture is constructed by connecting multiple RDCCUs in series. Figure 4 .

[0028] Furthermore, the core of the residual dilated causal convolution unit is the dilated causal convolution, and its structure is as follows Figure 5 As shown, the mathematical model is expressed as: ; In the formula, x and yRepresent the input and output of the convolution respectively, s is the current time step, f The size is K The convolution kernel, i is the convolution operation index, m is the expansion factor, represents time series data, x s is the time step s The input signal value at .

[0029] Causal convolution uses only the current and historical input data for calculation, which not only meets the needs of power load forecasting, but also prevents future information leakage. m (doubled per layer), the model can expand the receptive field while keeping the depth unchanged, improving the training effect. At the same time, the input of the previous layer is periodically sampled. m The calculation is: ; In the formula, L The computed output length for the convolution operation.

[0030] When the kernel size is fixed at 3, the dilation factor m Increasing from 1 to 4 will extend the effective history length of the output layer neurons from 3 to 15. To maintain sequence integrity, the output of each layer is aligned to the input sequence length by zero padding.

[0031] Furthermore, S6 is specifically as follows: input the test set into charging load prediction models with different numbers of layers, and evaluate the prediction performance of each model by calculating MAE and RMSE. The calculation formula is as follows: ; In the formula, N is the number of tests, i ∈[1, N ], y i For the i The predicted load of the test, y fi For the i The actual load of the test; The number of layers corresponding to the minimum values ​​of MAE and RMSE is selected as the optimal parameter of the multi-cascade LSTM-BDL model.

[0032] like Figure 6As shown in the figure, the mean absolute error (MAE) and root mean square error (RMSE) test results of different numbers of layers show that the 5-layer LSTM-BDL model has the best performance. When the network depth increases from 2 layers to 5 layers, the prediction performance gradually improves; but continuing to increase the number of layers leads to a decrease in the root mean square error (RMSE) and mean absolute error (MAE) values. This shows that moderately increasing the network depth and parameters can significantly improve the prediction performance, but excessive increase may cause overfitting due to parameter redundancy and reduced data diversity. At the same time, more layers and training parameters also mean longer training time.

[0033] Furthermore, in the embodiment of the present invention, the charging load data of a northern city from January 1, 2023 to March 1, 2024 is used for testing. The city's 38 public DC charging stations are distributed in residential areas (8), commercial areas (12), entertainment areas (11) and other areas (7), with a total of 298 charging poles, and a single pole capacity of 60kW. The data set contains active power, transaction power and charging period information, with a sampling interval of 15 minutes.

[0034] The experiment uses meteorological data such as temperature, humidity, precipitation, visibility and wind speed provided by China Weather Information Network, aligns them with load data through spline interpolation, and introduces time series features such as date type and season as additional information. The experiment is implemented based on the Pytorch framework in the Python 3.6 environment, and the hardware configuration is Inteli9-13980K CPU, NVIDIA RTX4060 GPU and 16GB RAM.

[0035] This example uses three indicators, namely, mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE), to evaluate model performance. MAPE reflects model accuracy by the percentage of predicted error to actual value: ; MAPE provides an intuitive relative error measure, but can lead to larger errors when the true value is close to zero. The lower the values ​​of these evaluation indicators, the higher the prediction accuracy of the model.

[0036] Winkler and Pinball are commonly used indicators for evaluating the performance of probability forecasting: Winkler reflects the width and coverage of the prediction interval, and Pinball measures the accuracy and reliability of the model. The specific expression is as follows: ; ; In the formula, q represents the quantile (Pinball takes 0.5), M t ,m t Respectively represent the time step t The upper and lower limits of the predicted load, represents a constant used to adjust the score, , Respectively represent the time step t The actual load value and q The predicted load value of the quantile, α is the score adjustment constant (Winkler takes 0.1).

[0037] This example compares the MCBDL-TCN model with the ANN and LSTM models, which use the same training configuration: mean square error loss function, Adam optimizer, batch size 672, and learning rate 0.002. The dataset is divided into training set, validation set, and test set in a ratio of 8:1:1.

[0038] The load forecasting results of the MCBDL-TCN model, ANN model and LSTM model are shown in Figure 7 , Figure 8 and Fig. 9 As shown. Although the predicted values ​​of the three models deviate from the true values, the ANN model basically reflects the actual load change trend. In contrast, the MCBDL-TCN model of this embodiment performs more stably, and the predicted values ​​are closer to the actual values, especially maintaining a high accuracy in the rising stage of the load curve. The LSTM and ANN models show obvious lag at the load peak. By comparison, it can be seen that the MCBDL-TCN model can simultaneously learn the long-term change law and fluctuation characteristics of electric vehicle load, thereby improving the prediction accuracy.

[0039] In order to evaluate the prediction performance of the MCBDL-TCN network, this embodiment uses a test data set to predict the charging load in four seasons and compares it with the existing point prediction and probability prediction methods. The comparison results are shown in Table 1. The RMSE, MAE and MAPE indicators of each model are as follows: Fig.10 , Fig.11 and Fig.12 As shown: Table 1 Comparison of prediction errors of models in different seasons From the perspective of RMSE, MAE, and MAPE (lower values ​​indicate higher prediction accuracy), the MCBDL-TCN algorithm outperforms the LSTM and ANN models in all seasons. Specifically, compared with LSTM, the MCBDL-TCN algorithm reduces the three indicators by 52.7%, 50.0%, and 58.7% in spring, 52.0%, 52.4%, and 55.7% in summer, 58.8%, 59.8%, and 60.7% in autumn, and 55.0%, 54.0%, and 57.2% in winter. These significant improvements show that the MCBDL-TCN model is the best choice for handling uncertainty in load forecasting of electric vehicle charging stations by effectively capturing global and local time series dependencies.

[0040] This embodiment performs probabilistic prediction of charging load on working days and non-working days, and compares it with a prediction method based on quantile regression (QR). Fig.13 Table 2 shows the bar graphs and average losses of the two methods. Fig.14 and Table 3 shows their Winkler and Pinball probability evaluation metrics.

[0041] Table 2 Performance of the two algorithms Table 3 Winkler and Pinball indicators of the two algorithms It can be seen that the probability evaluation indicators of the model proposed in the embodiment of the present invention are obviously better than those of the QR method.

[0042] The electric vehicle charging load prediction system based on the MCBDL-TCN model proposed in this embodiment includes the following five modules: a data acquisition module is responsible for collecting charging station demand data and local meteorological data; a data processing module processes the charging demand data; a charging load prediction module is constructed based on multiple cascaded Bayesian deep learning and time convolution models; a model evaluation module evaluates the prediction performance; and a model optimization module performs model optimization.

[0043] The electric vehicle charging load prediction system based on the MCBDL-TCN model proposed in this embodiment includes five modules: a data acquisition module is responsible for collecting charging station demand data and local meteorological data; a data processing module processes the charging demand data; a charging load prediction module is built based on multiple cascaded Bayesian deep learning and time convolution models; a model evaluation module evaluates the prediction performance; and a model optimization module performs model optimization.

[0044] Furthermore, the charging load prediction model consists of three parts: a multi-cascade Bayesian deep learning model formed by cascading multiple LSTM-BDLs, a temporal convolution model, and a Linear layer.

[0045] Furthermore, the temporal convolution model is composed of multiple residual dilated causal convolution units connected in series, each of which contains two paths: the main path is composed of two groups of identical structures connected in series, each of which is a dilated causal convolution factor, a Weight Norm layer, a ReLu activation function, and a Dropout layer; the residual path is a 1×1 convolution layer, and finally the output of the second Dropout layer of the main path is connected to the output of the 1×1 convolution layer of the residual path In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0046] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting electric vehicle charging load based on MCBDL-TCN model, characterized in that: The following steps are involved: S1. Obtain charging demand data of electric vehicle charging stations and corresponding meteorological data; S2, converting the charging demand data and meteorological data of electric vehicle charging stations into time series data, and dividing them into training set, validation set and test set; S3. Construct a charging load prediction model based on a multi-cascade Bayesian deep learning model and a time convolution model; The MCBDL architecture realizes progressive extraction of data features by connecting multiple LSTM-BDL units in series. First, LSTM-BDL is used to process time series input to obtain uncertainty features, and then the uncertainty features are passed to the TCN module for in-depth analysis. This combination design fully utilizes the advantages of the two networks, enhances the ability to process complex time series data, and improves the capture of global dependencies and multi-scale features. Finally, the output information of MCBDL and TCN is integrated through the fully connected layer to generate prediction results. The core component of TCN is RDCCU, which contains two rounds of causal convolution components with the same expansion, Weight Norm layer, activation function, Dropout layer and input residual connection. A multi-layer TCN network architecture is constructed by connecting multiple RDCCUs in series. S4. Optimize the weight W and bias b of LSTM-BDL in combination with Bayesian theory, use standard normal distribution as the prior distribution of weight W and bias b, determine the posterior distribution of weight W and bias b according to Bayesian theory, and process the posterior distribution through variational inference method to obtain the optimal distribution of weight W and bias; S5. Training a charging load prediction model based on the training set and the validation set; S6. Using the test data set, the MAE and RMSE indicators are calculated for MCBDL models of different depths, and the optimal number of layers of the LSTM-BDL model is determined based on the minimum value of MAE and RMSE; S7. Collect historical operation data and environmental parameters of the target charging station, convert this information into a standard time series format, and input it into the trained charging load prediction model to obtain the final load prediction value.

2. The electric vehicle charging load prediction method based on the MCBDL-TCN model according to claim 1 is characterized in that: S1 collects the operating parameters of the charging equipment at each station, including actual power, transaction power and charging period, and records the power value every 15 minutes; simultaneously obtains the corresponding meteorological data including precipitation, temperature, air humidity, visibility and wind speed, and implements spline interpolation processing on the meteorological data.

3. The electric vehicle charging load prediction method based on the MCBDL-TCN model according to claim 1 is characterized in that: S3 is as follows: After the feature information extracted by MCBDL is passed into TCN, each layer of LSTM-BDL modules continuously analyzes the uncertainty in the data. This design effectively reduces the error accumulation in complex data processing and improves the system's modeling accuracy and robustness for time series variables. The cascade function of the charging load prediction model is expressed as follows: ； In the formula, is the output of LSTM-BDL, l is the number of cycles of LSTM-BDL, n is the multiplicity of the cascade, for l- 1 layer at the current time step t The hidden state output of for l The layer at the current time step t The hidden state output is -1, For l Layer at the previous time step t -1 memory cell status; The original intention of the design of fusing LSTM-BDL and TCN is to integrate their respective advantages: the former's expertise in sequence modeling and uncertainty processing, and the latter's high efficiency in temporal dependency extraction. The MCBDL-TCN architecture constructed in this way enhances the ability to analyze complex data, especially when processing the dynamic characteristics of load data, ensuring the reliability of the prediction results.

4. The electric vehicle charging load prediction method based on the MCBDL-TCN model according to claim 3 is characterized in that: The calculation function of LSTM-BDL is: ； ； In the formula, is the current input, σ represents the sigmoid function, and Represent the weight and bias of LSTM-BDL respectively, f t is the forgotten value, o t is the output value, c t For the memory unit, is a candidate memory unit, i t To update the value, h t-1 is the hidden state at the previous moment, h t is the hidden state at the current moment.

5. The electric vehicle charging load prediction method based on the MCBDL-TCN model according to claim 3 is characterized in that: The temporal convolution model consists of multiple residual dilated causal convolution units connected in series, and its unit structure is: ； In the formula, o Represents the result after activation function processing, using Artivation Activation function, x represents the output from the previous layer, F(x ) represents the residual function.

6. The electric vehicle charging load prediction method based on the MCBDL-TCN model according to claim 5 is characterized in that: The core of the residual dilated causal convolution unit is the dilated causal convolution, and the mathematical model is expressed as: ； In the formula, x and y Represent the input and output of the convolution respectively, s is the current time step, f The size is K The convolution kernel, i is the convolution operation index, m is the expansion factor, represents time series data, x s is the time step s The input signal value at ; Causal convolution uses only the input data of the current and historical moments for calculation, which not only meets the needs of power load forecasting, but also prevents future information leakage. m , which can expand the receptive field while keeping the depth unchanged and improve the training effect; at the same time, the front layer input is periodically sampled. m The calculation is: ； In the formula, L The computed output length for the convolution operation.

7. The electric vehicle charging load prediction method based on the MCBDL-TCN model according to claim 1 is characterized in that: S6 is as follows: the test set is input into charging load prediction models with different numbers of layers, and the prediction performance of each model is evaluated by calculating MAE and RMSE. The calculation formula is as follows: ； In the formula, N is the number of tests, i ∈[1, N ], y i For the i The predicted load of the test, y fi For the i The actual load of the test; The optimal number of layers of the LSTM-BDL model is determined based on the minimum values ​​of MAE and RMSE.

8. An electric vehicle charging load prediction system based on the MCBDL-TCN model, characterized in that: It includes the following five modules: the data acquisition module is responsible for collecting charging station demand data and local meteorological data; the data processing module processes the charging demand data; the charging load prediction module is built based on multiple cascaded Bayesian deep learning and time convolution models; the model evaluation module evaluates the prediction performance; and the model optimization module performs model optimization.

9. The electric vehicle charging load prediction system based on the MCBDL-TCN model according to claim 8 is characterized in that: The charging load prediction model consists of three parts: a multi-cascade Bayesian deep learning model formed by cascading multiple LSTM-BDLs, a temporal convolution model, and a Linear layer.

10. The electric vehicle charging load prediction system based on the MCBDL-TCN model according to claim 9 is characterized in that: The temporal convolution model is composed of multiple residual dilated causal convolution units connected in series, each of which contains two paths: the main path is composed of two groups of identical structures connected in series, each of which is an expanded causal convolution factor, a Weight Norm layer, a Relu activation function, and a Dropout layer; the residual path is a 1×1 convolution layer, and finally the output of the second Dropout layer of the main path is connected to the output of the 1×1 convolution layer of the residual path.

Images