Electric vehicle load prediction method based on knowledge graph attention network and deep learning

Abstract

The invention discloses an electric vehicle load prediction method based on a knowledge graph attention network and deep learning, and the method improves the prediction precision through integrating domain knowledge into a KAN structure and combining the modeling of the CNN on the space-time dependence. In addition, in consideration of the fact that a plurality of charging stations are possibly managed by different operators, a cloud edge collaborative architecture is constructed, data privacy protection is ensured, and plug and play prediction of new charging stations is supported. Simulation experiments show that compared with a traditional method, KAN-CNN has higher prediction precision and lower calculation time and resource consumption, can effectively process a multi-charging-station scene, and has a good engineering application prospect.

Description

Technical Field

[0001] The present invention relates to charging load prediction technology, and in particular to an electric vehicle load prediction method based on knowledge graph attention network and deep learning. Background Art

[0002] With the rapid popularization of electric vehicles, their impact on active distribution networks has become increasingly significant. This impact lies not only in the increasing load scale driven by the growth in the number of electric vehicles, but also in the peak load superposition effect of charging behavior, which leads to operational challenges such as local voltage fluctuations and line overloads. In addition, the randomness and spatiotemporal variability of electric vehicle charging and discharging patterns overturn the traditional assumption of load stability and put forward new requirements for existing power dispatching and optimization strategies. However, this uncertainty also highlights the potential of electric vehicles as flexible loads and distributed energy storage resources, providing new opportunities for demand response and active dispatching. Therefore, accurate prediction of electric vehicle load behavior is not only the basis for ensuring the safe operation of distribution networks, but also crucial for optimizing resource allocation, improving demand response efficiency, and promoting the development of future smart grids.

[0003] Traditional EV load forecasting relies primarily on model-driven approaches that utilize mathematical frameworks and predefined assumptions about charging patterns. For example, Monte Carlo methods have been used to simulate the spatiotemporal distribution of EV load demand, while Gaussian mixture models and decision trees have been used to predict the load at charging stations. In addition, probability-based models have been developed that incorporate detailed characteristics of different vehicle types and destinations to improve forecast accuracy. While these approaches perform well under predictable conditions, they often have difficulty addressing the dynamic, nonlinear, and stochastic nature of real-world EV charging behavior. As EV adoption becomes more complex and diverse, the limitations of model-driven approaches become apparent, highlighting the need for adaptive data-driven approaches to effectively handle this uncertainty.

[0004] The development of machine learning (ML) technology has promoted the great application prospects of data-driven methods in electric vehicle load forecasting. Among them, multi-layer perceptron (MLP) is widely used due to its flexibility and generalization ability. Various techniques (such as neural networks, deep learning networks, GRU-RNN and Transformer) have been applied in the existing technology to predict electric vehicle load based on charging demand and user behavior constraints. However, the inherent structure of MLP has limitations in capturing the temporal dependencies and contextual relationships in time series data, resulting in poor performance in modeling dynamic and complex electric vehicle load patterns. To address these challenges, many researchers have combined multiple data-driven models. For example, the reference (C.Li, et al. Prediction of EV charging load using two-stage time series decompositionand DeepBiLSTM model. IEEE Access, 2023, 11, pp.72925-72941. DOI: 10.1109 / ACCESS.2023.3294273.) uses variational mode decomposition (VMD) for load decomposition and then applies Bi-LSTM for prediction, while the reference (F.Mohammad, et al. Energy demand load forecasting for electricvehicle charging stations network based on convlstm and biconvlstmarchitectures. IEEE Access, 2023, 11, pp 67350-67369. DOI: 10.1109 / ACCESS.2023.3274657.) combines models such as ConvLSTM and BiConvLSTM architectures to improve prediction accuracy. Despite the improvement in prediction accuracy, these methods usually require large computational resources and are highly sensitive to parameter tuning, limiting their practical applicability. This problem stems from the inherent structure of MLP, which remains difficult to optimize for these challenges, highlighting the urgent need for a new prediction framework that can simplify the structure without compromising the prediction performance, thereby improving the engineering feasibility of EV load prediction.

[0005] like Figure 1As shown in the figure, KAN and MLP are essentially different. KAN is based on the Kolmogorov-Arnold representation theorem, which states that any continuous function can be approximated as the sum of simpler single-variable functions. This allows KAN to focus on the effective approximation of complex multidimensional functions. MLP: does not rely on any specific theoretical basis. It learns through neuron layers, each of which performs a nonlinear transformation of the input data and relies on an empirical training process.

[0006] Therefore, KAN can clearly decompose the multi-dimensional input function into single variable components, thereby reducing complexity. This structured approach has obvious advantages in electric vehicle load forecasting. MLP directly approximates the entire complex function, which is less efficient in high-dimensional or highly nonlinear environments. This defect will directly lead to forecasting failure when there are short-term fluctuations in electric vehicle load.

[0007] In addition, due to its function decomposition, KAN requires fewer parameters, converges faster, and is generally more computationally efficient. This is very useful in large-scale forecasting applications such as electric vehicle load forecasting, where computing resources are usually limited. MLPs often require more parameters and more iterations to converge, especially when dealing with high-dimensional data, which increases the computational cost. At the same time, as the number of parameters increases, the cost of MLP in parameter tuning cannot be ignored.

[0008] Overall, KANs provide a more structured, efficient, and powerful framework for modeling complex systems, while MLPs, while flexible and widely applicable, may require more resources and extensive fine-tuning to achieve similar performance in challenging environments. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide an electric vehicle load forecasting method based on knowledge graph attention network and deep learning to improve the accuracy of electric vehicle load forecasting in view of the shortcomings of the existing technology.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for electric vehicle load forecasting based on knowledge graph attention network and deep learning, comprising the following steps:

[0011] Acquire historical data of the charging station, and pre-process the historical data of the charging station;

[0012] The preprocessed data is used as an input of a KAN-CNN network, and the KAN-CNN network is trained to obtain a load forecasting model;

[0013] The KAN-CNN network includes a Kolmogorov-Arnold network and a convolutional neural network; the Kolmogorov-Arnold network is connected in series with the convolutional neural network, and the output of the convolutional neural network is connected to a fully connected layer.

[0014] The specific implementation process of preprocessing the charging station historical data includes:

[0015] Fill in missing data;

[0016] Normalize the padded data;

[0017] The normalized data is denormalized to obtain the preprocessed data.

[0018] The output of the fully connected layer is expressed as:

[0019]

[0020] Among them, F k is the kth feature map, is the activation function, v k and c are weight and bias respectively, To predict the results, m represents the dimension of the input features, that is, the number of features used.

[0021] Among them, w ki is the convolution kernel parameter, Z i is the i-th eigenvector output by the Kolmogorov-Arnold network, b k represents the bias term, and n is the dimension of the feature vector output by the Kolmogorov-Arnold network.

[0022] The method of the present invention further comprises:

[0023] The charging station data is used as input of the load prediction model to obtain a load prediction value.

[0024] As an inventive concept, the present invention also provides an electric vehicle load forecasting system comprising:

[0025] The local server is used to obtain the historical data of the charging station, pre-process the historical data of the charging station, use the pre-processed data as the input of the KAN-CNN network, train the KAN-CNN network, and upload the trained weight parameters to the cloud server; update the weight parameters of the KAN-CNN network according to the aggregated parameters sent by the cloud server to obtain the trained load forecasting model;

[0026] The cloud server is used to obtain the trained weight parameters, aggregate the trained weight parameters obtained each time, and send the aggregated parameters to the local server.

[0027] As an inventive concept, the present invention also provides an electric vehicle load forecasting system, comprising a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the above method.

[0028] As an inventive concept, the present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon; the computer program / instruction implements the steps of the above method when executed by a processor.

[0029] As an inventive concept, the present invention also provides a computer program product, including a computer program / instruction; when the computer program / instruction is executed by a processor, the steps of the above method are implemented.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. The KAN-CNN model of the present invention outperforms the current mainstream Bi-LSTM and Bi-GRU models in evaluation indicators such as RMSE, MAPE and R2. For example, KAN-CNN achieves lower RMSE and MAPE in all regions, and also achieves higher R2 values, indicating that its prediction results are more consistent with the actual load trend;

[0032] 2. Compared with Bi-LSTM and Bi-GRU, KAN-CNN significantly reduces the consumption of computing resources and execution time. In the experiment, the calculation time of KAN-CNN was reduced by about 50% compared with other methods, making it more suitable for engineering application scenarios such as real-time prediction;

[0033] 3. KAN-CNN combines the advantages of Kolmogorov-Arnold network (KAN) and convolutional neural network (CNN), effectively integrating domain knowledge with temporal and spatial feature extraction capabilities. This hybrid model not only maintains high prediction accuracy, but also significantly simplifies the model parameter adjustment process, ensuring scalability and reliability in complex and dynamic environments (such as active distribution networks);

[0034] 4. As the first application of KAN-CNN structure in power system load forecasting, this paper verifies its feasibility and engineering practical value. The model can also be effectively extended to other time series forecasting tasks in the energy field, such as renewable energy power generation forecasting or power grid stability analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1The difference between KAN and MLP;

[0036] Figure 2 This is a prediction framework diagram of an embodiment of the present invention;

[0037] Figure 3 The cloud-edge collaborative multi-charging station load prediction architecture of the embodiment of the present invention;

[0038] Figure 4 This is the KAN-CNN prediction result of the embodiment of the present invention;

[0039] Figure 5 Prediction result for Bi-LSTM;

[0040] Figure 6 This is the Bi-GRU prediction result. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.

[0042] Example 1

[0043] The embodiment of the present invention provides an electric vehicle load forecasting method, focusing on the integration of Kolmogorov-Arnold network (KAN) and convolutional neural network (CNN). Its essence is: decomposing data through KAN, and then extracting features with CNN to achieve efficient load forecasting. It includes the specific implementation of KAN components, CNN components and their combination.

[0044] 1. KAN Components - Decomposition Input Function

[0045] According to the Kolmogorov-Arnold representation theorem, any continuous multidimensional function can be represented as a weighted combination of several one-dimensional functions:

[0046]

[0047] Among them, g i (.) and h ij (.) are represented as one-dimensional functions respectively.

[0048] Then, KAN is used to decompose the periodicity and randomness characteristics of the charging load, that is, to map the original data X into several low-dimensional features:

[0049] Z=φ(X) (4)

[0050] Where φ(.) represents the decomposition function of KAN, and Z is the extracted feature vector.

[0051] After KAN decomposition, the load of the electric vehicle charging station is decomposed into several one-dimensional features, reducing the impact of feature coupling. In addition, the periodic characteristics of the load can be explicitly modeled to improve the prediction ability.

[0052] 2. CNN feature extraction stage - prediction

[0053] The feature vector Z after KAN decomposition is input into CNN to capture the local features of the time series. The core calculation formula of CNN is:

[0054]

[0055] Among them, F k is the kth feature map; w ki is the convolution kernel parameter; is the activation function.

[0056] The temporal feature output extracted by CNN is:

[0057] F={F1,F2,…,F m} (6)

[0058] These characteristics describe local dependencies of the load series (such as short-term fluctuations).

[0059] 3. KAN-CNN fusion and prediction

[0060] The decomposition features of KAN and the temporal features of CNN are fused and input into the fully connected layer:

[0061]

[0062] Among them, v k and c are weight and bias respectively; For the prediction results.

[0063] The evaluation of the prediction results mainly includes the root mean square error (RMSE), the mean absolute percentage error (MAPE), and the coefficient of determination (R-squared, R2):

[0064] 1. Root Mean Squared Error (RMSE)

[0065] The root mean square error (RMSE) is the square root of the mean square error (MSE). It is more sensitive to larger errors and can therefore reflect significant deviations in the prediction results. The smaller the RMSE value, the lower the prediction error of the model and the better the prediction effect. In addition, since its unit is consistent with the original data, it has good interpretability. The formula for RMSE is as follows:

[0066]

[0067] Among them, y i is the actual value is the predicted value, and N is the number of samples.

[0068] 2. Mean Absolute Percentage Error (MAPE)

[0069] Mean absolute percentage error (MAPE) quantifies the prediction error in percentage form. It is a standardized indicator that facilitates comparison of prediction performance under different load ranges. The smaller its value, the better the prediction effect of the model. The formula of MAPE is as follows:

[0070]

[0071] MAPE can more intuitively reflect the deviation between the predicted results and the actual values ​​by introducing the concept of relative error.

[0072] 3. Coefficient of determination (R-squared, R 2 )

[0073] Coefficient of determination (R 2 ) is an important indicator to measure the model's ability to explain data changes. Its value range is [0,1]. The closer it is to 1, the better the model fits the data changes.

[0074] The formula is as follows:

[0075]

[0076] Among them, y i is the actual value, is the predicted value, is the average of the actual values; the numerator represents the residual sum of squares, and the denominator represents the total sum of squares. 2 The value indicates that the model can better explain the trend of data changes.

[0077] The above three indicators can be used to comprehensively evaluate the performance of the prediction model. Among them, RMSE is used to reflect the absolute error of the prediction result, MAPE is used to compare the relative errors of different load ranges, and R 2The model's ability to explain the overall data change trend is measured. These indicators provide a scientific basis for the subsequent performance comparison of comprehensive load forecasting of multiple charging stations.

[0078] Considering that there are multiple charging stations in the entire area, and the above charging stations may belong to different operators, uploading historical data uniformly for prediction will infringe the interests of different operators. In addition, plug-and-play prediction is required for newly deployed charging stations, so a cloud-edge collaborative architecture for comprehensive prediction of multiple charging stations is constructed here. Figure 3 shown.

[0079] In this framework, each participating charging station holds the original user data collected by itself, and there are barriers between these data, that is, they cannot directly exchange original data. This barrier includes data exchange between charging stations and between charging stations and cloud servers. This model ensures that no party can access the data information of other participants, so that the data privacy of all parties is protected.

[0080] In the initial stage, the server uses a public dataset in a specific field (without considering privacy issues) for pre-training to initialize the cloud model parameters and distribute the model parameters to local participants. Each participant deploys the model locally in advance and uses the received model parameters to update its own model parameters, and then trains the model using local data. The local training process is based on the KAN-CNN network model, and the goal of all charging stations is to train a shared prediction model.

[0081] After training is completed, in order to reduce the risk of data leakage, participants upload the trained weight parameters instead of the gradients to the cloud server. This method can effectively prevent attacks on model gradients and potential threats of data theft. In the framework, each local model will fuse and return model parameters during the optimization process. The fusion process involves the aggregation of parameters of multiple local models, while the return process is to update the parameters of each local model. After a certain number of iterations, the global objective function will reach convergence, thus completing the training of the model.

[0082] This experiment uses the historical load data of all charging stations in a certain area from 00:00 on September 1, 2023 to 24:00 on August 31, 2024 for modeling and prediction. The experimental data is collected at a time interval of 15 minutes, and a total of 35,136 time points are collected, covering the charging pile loads in seven areas of AG. The number of regional charging piles ranges from 2,487 to 12,986, which fully reflects the load characteristics of different regions. The experiment is divided into a training set (September 2023 to July 2024, accounting for 90%), a validation set (August 1 to 15, 2024, accounting for 5%), and a test set (August 16 to 31, 2024, accounting for 5%). The model input is the load data of the past 120 hours, and the prediction target is the short-term load in the next 24 hours.

[0083] All model training and testing were completed on a desktop computer with the following configuration: operating system Ubuntu 22.04, processor AMD Ryzen 9 7950X 16-core 32-thread CPU, graphics card NVIDIA GeForce RTX 4090 GPU, 24GB video memory, 64GB DDR5 memory, and 2TB NVMe SSD storage. The deep learning framework uses PyTorch 2.0, and data preprocessing and experimental analysis are completed in Python 3.10 environment. This hardware platform ensures the efficiency of model training and the accuracy of prediction results.

[0084] Figure 4 The data from 0:00 on August 1 to 24:00 on August 5 is used to predict the charging station load data results on August 6. At the same time, in order to verify the superiority of the algorithm described in the embodiment of the present invention, the more advanced Bi-LSTM and Bi-GRU algorithms are used to verify the same data, and the results are shown in the figure. The statistical indicators described in the embodiment of the present invention are used to compare the prediction accuracy of different algorithms as shown in Table 1.

[0085] Table 1 Comparison statistics of prediction results of different algorithms

[0086]

[0087] Based on the statistical results in Table 1, it can be seen that the KAN-CNN model is superior to Bi-LSTM and Bi-GRU in all evaluation indicators, as shown below:

[0088] Root Mean Square Error (RMSE):

[0089] In area A, the RMSE of KAN-CNN is 530.220MW, which is closer to the actual value of 525.260MW than 505.260MW of Bi-LSTM and 545.290MW of Bi-GRU, reflecting higher prediction accuracy. RMSE (root mean square error), as an important indicator for measuring prediction deviation, not only reflects the average error between the predicted value and the actual value, but also gives higher weight to larger errors. Therefore, although the RMSE of KAN-CNN is slightly higher than that of Bi-LSTM, the trend of being closer to the actual value shows that it has a stronger ability to capture overall load fluctuations. Although Bi-LSTM seems to be closer to the actual value in terms of value, there may be a risk of deviation amplification in data over a longer period of time or in other areas, which can also be reflected in the results of other areas. The RMSE of Bi-GRU is significantly higher than the actual value, indicating that there is a larger error in its prediction results and the model is less sensitive to load fluctuations in specific areas.

[0090] In region D, the RMSE of KAN-CNN is 148.183MW, which is significantly better than Bi-LSTM (171.086MW) and Bi-GRU (137.006MW). The reason why KAN-CNN performs well in this region may be that its convolutional neural network architecture can better capture the local patterns and short-term fluctuation characteristics in the load data, while Bi-LSTM and Bi-GRU rely on recursive processing of sequence information and may not capture the local features in the data enough. In addition, although the RMSE of Bi-GRU is numerically slightly lower than that of KAN-CNN, considering the performance of other evaluation indicators (such as MAPE and R2), the overall prediction ability of KAN-CNN is still more advantageous. Especially in the case of large load fluctuations, KAN-CNN can more stably generate predictions close to the actual values, avoiding over-smoothing or excessive sensitivity to outliers.

[0091] The difference in RMSE reflects the adaptability and performance characteristics of different models under different load environments from an important perspective. The RMSE of KAN-CNN is close to the actual value in both areas A and D, indicating that it is robust in overall error control and can effectively avoid the occurrence of large errors; although Bi-LSTM performs well in some areas, it may not be adaptable enough to complex load patterns; although Bi-GRU has a slight advantage in a single indicator (such as lower RMSE), after comprehensive evaluation, its model's ability to capture data characteristics is slightly insufficient. This further verifies the advantages of KAN-CNN in charging station load forecasting, and its potential value in engineering applications is particularly significant.

[0092] Mean Absolute Percent Error (MAPE):

[0093] The MAPE value in area D shows that KAN-CNN is 6.583%, which is significantly lower than Bi-LSTM's 8.965% and slightly better than Bi-GRU's 6.963%. Further analysis shows that the lower MAPE value means that KAN-CNN has a relatively small deviation from the actual value during the prediction process, and its error margin is smaller and more stable. This is especially important because MAPE quantifies the error size in percentage form, which can more intuitively reflect the accuracy of the prediction results under different load ranges.

[0094] For example, the charging station load in area D has a certain volatility, and the load value changes greatly. In this case, the lower MAPE shows that the KAN-CNN model can not only accurately capture the overall change trend of the load, but also adapt well to the local fluctuation characteristics of the load, further verifying its applicability and robustness in complex load environments.

[0095] A smaller MAPE value also means that in practical applications such as load scheduling or energy optimization, the error in KAN-CNN prediction will lead to smaller deviations in resource allocation, which helps to more accurately plan the capacity and energy distribution of charging facilities. Therefore, from the perspective of practical engineering, KAN-CNN not only improves the reliability of predictions, but also provides stronger support for the decision-making optimization of operators. It is further proved that KAN-CNN has better error performance under different load ranges.

[0096] Coefficient of determination (R2):

[0097] The coefficient of determination (R2) is an important statistical indicator for evaluating model performance. Its value range is (0, 1). The closer it is to 1, the stronger the model's ability to explain data changes and the closer the predicted results are to the actual values. The calculation of R2 is based on the proportional relationship between the total error (the deviation between the actual value and the mean) and the residual (the deviation between the actual value and the predicted value). The larger the R2 value, the higher the model's explanation of the changes in the target variable.

[0098] In region C, the R2 of KAN-CNN reaches 0.983, compared with 0.942 for Bi-LSTM and 0.967 for Bi-GRU. From the numerical difference, it can be seen that the KAN-CNN model has a significantly better ability to fit the data trend than other models, which indicates that it can more accurately capture the overall change law and local characteristics of the load in this region. The R2 value of KAN-CNN is close to 1 in all regions. For example, in region C, the R2 of KAN-CNN is 0.983, which is significantly better than 0.942 for Bi-LSTM and 0.967 for Bi-GRU, indicating that KAN-CNN has the strongest ability to explain the trend of data changes.

[0099] The R2 value of Bi-LSTM (0.942) is low, indicating that its ability to explain data changes is limited. This may be because Bi-LSTM relies on time-step recursive processing of sequence data. When faced with complex load fluctuations, it is difficult to effectively extract the key features of the data, causing the prediction results to deviate from the actual value. The R2 value of Bi-GRU (0.967) is better than Bi-LSTM, indicating that its ability to capture load data has improved, but there is still a gap compared with KAN-CNN. This may be related to the gating mechanism of Bi-GRU. It has a better ability to handle long-term dependencies, but may not be as efficient as KAN-CNN in extracting short-term local features.

[0100] In contrast, the R2 of KAN-CNN is 0.983, which is close to 1, indicating that the error between its predicted value and the actual value is very small, which can effectively explain the trend of data changes. KAN-CNN captures the local pattern of data through convolution operations, and combines the pooling layer to reduce unnecessary noise, so that it can show higher robustness and accuracy in complex load scenarios. In summary, the high prediction accuracy of the KAN-CNN model can be attributed to its convolutional network structure, which can more effectively extract data features, especially in scenarios with large load fluctuations in different regions.

[0101] KAN-CNN achieved an R2 value of 0.983 in region C, indicating that it has wider applicability and reliability in practical engineering applications. Specifically: In large-scale charging station load forecasting, data fluctuations are often more drastic, and the load distribution differences between regions are large. The high R2 value means that KAN-CNN can accurately capture the unique load characteristics of each region, providing a reliable basis for the dynamic scheduling and optimization of charging piles. The high R2 value also shows that KAN-CNN has a strong ability to handle noisy data. In actual scenarios, load data may fluctuate greatly due to external factors (such as weather and holidays), and KAN-CNN can model these influences without losing important information, ensuring the stability and credibility of the prediction results. The R2 value of KAN-CNN in region C reached 0.983, which is not only significantly better than Bi-LSTM (0.942) and Bi-GRU (0.967), but also further verifies its advantage in explaining data change trends. Through the characteristics of convolutional neural networks, KAN-CNN can effectively balance the overall trend and local fluctuations, significantly improving the adaptability and robustness of the model. This performance is crucial for load forecasting in practical application scenarios, because a higher R2 value means that the model can provide more accurate and stable forecasting support for the planning and management of charging stations.

[0102] In terms of computational efficiency, the KAN-CNN model also shows obvious advantages:

[0103] Data analysis and comparison:

[0104] Taking area A as an example, the calculation time of KAN-CNN is 15.6 seconds, which is 57.5% and 59.7% shorter than Bi-LSTM (36.7 seconds) and Bi-GRU (38.6 seconds); in area C, the calculation time of KAN-CNN is 14.5 seconds, which is 55.4% and 60.3% shorter than Bi-LSTM and Bi-GRU respectively. This significant time reduction shows that the advantage of KAN-CNN in computational efficiency is stable and universal, and is applicable to charging station load prediction tasks in different regions and of different scales.

[0105] Analysis of the internal mechanism of improving computing efficiency:

[0106] The improvement in KAN-CNN’s computational efficiency is due to the following two key factors.

[0107] Parameter optimization design: KAN-CNN uses a lighter parameter structure in network design, which significantly reduces the number of parameters in the calculation process compared to Bi-LSTM and Bi-GRU. Bi-LSTM and Bi-GRU need to update the hidden state at each time step. This recursive calculation feature causes the number of parameters and the amount of calculation to increase exponentially with the increase of time steps. KAN-CNN extracts data features through convolutional layers, does not rely on recursive operations of time steps, and has a fixed and relatively small parameter scale, which significantly reduces the computational complexity.

[0108] Parallel computing capabilities

[0109] One of the core advantages of convolution operations is the ability to parallelize computation. KAN-CNN can process multiple blocks of input data simultaneously, which greatly reduces computation time. However, due to the characteristics of their recurrent neural networks (RNNs), Bi-LSTM and Bi-GRU need to process data sequentially time-step by time, and cannot achieve true parallelization. This difference in architecture is the main reason why KAN-CNN leads in computational efficiency.

[0110] The improvement of KAN-CNN's computational efficiency is due to its optimized design of parameter quantity and the high efficiency of convolution operation. Compared with the complex process of Bi-LSTM and Bi-GRU that need to update the hidden state step by step, KAN-CNN can process input data in parallel, so it shows a significant time advantage in actual calculation.

[0111] In practical applications, the improvement in computational efficiency of KAN-CNN has brought significant engineering value:

[0112] Real-time requirements: Charging station load forecasting usually requires real-time processing of large amounts of data, such as high-frequency load change trend updates or abnormal load detection. KAN-CNN shortens the calculation time so that load forecasting can quickly respond to dynamic changes and meet real-time requirements. For example, in the prediction of areas A and C, KAN-CNN only takes about 15 seconds to complete the prediction, while Bi-LSTM and Bi-GRU take about 35 seconds or even longer. This time difference is particularly important in scenarios with high-frequency data updates.

[0113] Resource consumption and cost optimization: Improved computing efficiency also means less hardware resources are occupied. For example, with the same hardware configuration, KAN-CNN can save nearly 60% of computing time, which not only reduces computing costs, but also processes more tasks under the same resource conditions and improves the system's throughput. For large-scale charging station networks, this efficiency advantage helps reduce hardware investment and operating costs.

[0114] Scalability and universality: KAN-CNN’s high efficiency provides scalability for larger data sets or more complex prediction tasks. For example, as the number of charging stations increases or the load pattern becomes more complex, KAN-CNN can still complete predictions quickly and maintain high stability and reliability.

[0115] From the comprehensive analysis of prediction accuracy and computational efficiency, KAN-CNN has shown significant superiority in charging station load prediction in different regions. On the one hand, its prediction results are better than Bi-LSTM and Bi-GRU in terms of RMSE, MAPE and R2, and can more accurately reflect the regional load characteristics; on the other hand, its calculation time is significantly shortened, especially in areas A and C.

[0116] For example, in region D, KAN-CNN's MAPE is reduced by about 2.38% compared to Bi-LSTM, while the computation time is reduced by about 55%, which fully demonstrates the potential value of KAN-CNN in practical applications. Specifically, KAN-CNN is suitable for scenarios that require high-frequency real-time load forecasting, such as large-scale charging station load optimization and dynamic scheduling.

[0117] This study shows that KAN-CNN can not only significantly improve the prediction accuracy in charging station load forecasting, but also greatly reduce the demand for computing resources, meeting the dual requirements of efficiency and accuracy in engineering practice. This provides effective technical support for the optimization and load management of future smart charging networks, and also provides a reference for other time series forecasting tasks.

[0118] Example 2

[0119] Embodiment 2 of the present invention provides a system corresponding to the above-mentioned embodiment 1, including a memory, a processor and a computer program stored in the memory; the processor executes the computer program in the memory to implement the steps of the method in the above-mentioned embodiment 1.

[0120] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory.

[0121] In some other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors of various types, which are not limited herein.

[0122] Example 3

[0123] Embodiment 3 of the present invention provides a computer-readable storage medium corresponding to the above embodiment 1, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of the method of the above embodiment 1 are implemented.

[0124] Computer readable storage media can be tangible devices that hold and store instructions used by instruction execution devices. Computer readable storage media can be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any combination thereof.

[0125] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0126] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0127] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0128] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0129] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for electric vehicle load forecasting based on knowledge graph attention network and deep learning, characterized in that: The following steps are involved: Acquire historical data of the charging station, and pre-process the historical data of the charging station; The preprocessed data is used as an input of a KAN-CNN network, and the KAN-CNN network is trained to obtain a load forecasting model; The KAN-CNN network includes a Kolmogorov-Arnold network and a convolutional neural network; the Kolmogorov-Arnold network is connected in series with the convolutional neural network, and the output of the convolutional neural network is connected to a fully connected layer.

2. The electric vehicle load forecasting method based on knowledge graph attention network and deep learning according to claim 1 is characterized in that: The specific implementation process of preprocessing the charging station historical data includes: Fill in missing data; Normalize the filled data; The normalized data is denormalized to obtain the preprocessed data.

3. The electric vehicle load forecasting method based on knowledge graph attention network and deep learning according to claim 1 is characterized in that: The output of the fully connected layer is expressed as: Among them, F k is the kth feature map, is the activation function, v k and c are weight and bias respectively, To predict the results, m represents the dimension of the input features, that is, the number of features used.

4. The electric vehicle load forecasting method based on knowledge graph attention network and deep learning according to claim 3 is characterized in that: Among them, w ki is the convolution kernel parameter, Z i is the i-th eigenvector output by the Kolmogorov-Arnold network, b k represents the bias term, and n is the dimension of the feature vector output by the Kolmogorov-Arnold network.

5. The electric vehicle load forecasting method based on knowledge graph attention network and deep learning according to any one of claims 1 to 4, characterized in that: Also includes: The charging station data is used as input of the load prediction model to obtain a load prediction value.

6. An electric vehicle load forecasting system, characterized in that: include: The local server is used to obtain the historical data of the charging station, pre-process the historical data of the charging station, use the pre-processed data as the input of the KAN-CNN network, train the KAN-CNN network, and upload the trained weight parameters to the cloud server; update the weight parameters of the KAN-CNN network according to the aggregated parameters sent by the cloud server to obtain the trained load forecasting model; The cloud server is used to obtain the trained weight parameters, aggregate the trained weight parameters obtained each time, and send the aggregated parameters to the local server.

7. An electric vehicle load forecasting system, comprising a memory, a processor and a computer program stored in the memory; characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program / instruction stored thereon; characterized in that: When the computer program / instructions are executed by a processor, the steps of the method described in any one of claims 1 to 5 are implemented.

9. A computer program product comprising a computer program / instructions; characterized in that: When the computer program / instructions are executed by a processor, the steps of the method of any one of claims 1 to 5 are implemented.

Images