Electric vehicle charging pile health degree prediction method based on rkan-bilstm-sa

Abstract

The application is suitable for the technical field of charging piles, and provides a health degree prediction method for electric vehicle charging piles based on RKAN-BiLSTM-SA, which comprises the following steps: receiving time sequence input data, the input data comprising historical health degree grade data and key index features; inputting the input data into an RKAN layer for feature enhancement, comprising width expansion, nonlinear transformation and Dropout regularization processing in sequence; inputting the features output by the RKAN layer into a BiLSTM-SA layer, extracting time sequence features, and performing weighted fusion on the time sequence features through a self-attention mechanism; inputting the features output by the BiLSTM-SA layer into a full connection layer for nonlinear conversion and dimension reduction; and outputting a prediction result of future health degree of the charging pile. The application extracts nonlinear correlation features in the operation data of the charging pile through deep feature transformation and regularization constraint of the RKAN layer, and enhances the representation ability for complex health states.

Description

Technical Field

[0001] This invention relates to the field of charging pile technology, specifically to a method for predicting the health of electric vehicle charging piles based on RKAN-BiLSTM-SA. Background Technology

[0002] Traditional prediction models have inherent defects. Single LSTM (Long Short-Term Memory) or simple combination models have limitations when processing multi-dimensional operational data of charging piles. Ordinary recurrent neural networks are difficult to effectively capture long-term temporal dependencies. Conventional feedforward networks have insufficient ability to fit complex nonlinear relationships. Furthermore, the shallow coupling between modules leads to low feature transfer efficiency. These structural problems collectively restrict the expressive power of the model.

[0003] The health data of charging piles has significant inherent complexity. Operational indicators such as load rate and charging power exhibit strong multicollinearity, and fault modes show non-stationary and nonlinear evolution patterns. Key fault symptoms often show sparsity and lag in time series. These characteristics pose a huge challenge to accurate modeling.

[0004] Existing technologies first define the health status of charging piles by quantifying the remaining time before a fault occurs, and then construct a charging pile health status index system based on this. Next, they combine a KAN network with a Bidirectional Long Short-Term Memory (BiLSTM) network, introducing a regularization strategy into the KAN network to suppress overfitting and improve the model's generalization ability. Furthermore, they introduce a self-attention (SA) mechanism into the BiLSTM, which effectively amplifies key features through adaptive weighting, improving the model's ability to capture temporal dependencies. However, existing technologies, by integrating KAN networks and BiLSTM, and adding regularization and self-attention mechanisms, suffer from a large parameter scale, leading to high computational complexity. Moreover, they exhibit problems such as model overfitting and insufficient key feature capture ability, resulting in limited prediction accuracy and large prediction errors. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method for predicting the health of electric vehicle charging piles based on RKAN-BiLSTM-SA, so as to solve the problems in the background technology.

[0006] This invention is implemented as follows: a method for predicting the health of electric vehicle charging stations based on RKAN-BiLSTM-SA, the method comprising the following steps: Receive preprocessed time series input data, which includes historical health level data and seven key indicator features, including load rate, actual charging power, cumulative charging amount, failure rate, time to use, average duration of failure, and charging revenue. The input data is fed into the RKAN layer for feature enhancement, including sequential width expansion, non-linear transformation, and Dropout regularization. The features output from the RKAN layer are input into the BiLSTM-SA layer, and temporal features are extracted through a bidirectional long short-term memory network. The temporal features are then weighted and fused through a self-attention mechanism. The features output from the BiLSTM-SA layer are input into a fully connected layer for nonlinear transformation and dimensionality reduction. The output layer outputs the prediction results of the future health of the charging pile, and the results are the probability distribution of the health level at multiple future time steps.

[0007] As a further embodiment of the present invention, the RKAN layer includes a first KANLayer, a Dropout layer, a second KANLayer, and a third KANLayer connected in sequence; wherein, the first KANLayer expands the input feature dimension to 256 dimensions and uses the ReLU activation function; the second KANLayer compresses the feature dimension to 128 dimensions; and the third KANLayer outputs a feature vector of uniform dimension for use by the subsequent BiLSTM-SA layer.

[0008] As a further aspect of the present invention, the bidirectional long short-term memory network in the BiLSTM-SA layer includes a forward LSTM and a backward LSTM, which capture the temporal dependencies of the forward and backward directions, respectively; the number of hidden units in the bidirectional LSTM is 128, and the hidden state is output at each time step; the self-attention mechanism adopts a multi-head attention structure, including 4 attention heads, for extracting key information from different feature subspaces.

[0009] As a further aspect of the present invention, the specific steps of the self-attention mechanism include: Perform layer normalization on the feature sequence output by BiLSTM; Generate the query matrix, key matrix, and value matrix through linear transformation; Calculate the attention weights and perform Softmax normalization; The attention output for each time step is obtained by weighted summation of the value matrix.

[0010] As a further aspect of the present invention, the fully connected layer employs the ReLU activation function to compress the high-dimensional features output by the BiLSTM-SA layer to 64 dimensions and performs a nonlinear transformation.

[0011] As a further aspect of the present invention, the output layer uses the Softmax function to map the features output by the fully connected layer into a probability distribution of 5 health levels, representing the health status of the charging pile at each future time step.

[0012] As a further aspect of this invention, during the training of the RKAN-BiLSTM-SA model, the total loss function consists of cross-entropy loss and an L2 regularization term, specifically expressed as follows: , where L task The cross-entropy loss is λ, where λ is the regularization coefficient and W is the cross-entropy loss. l Let be the weight matrix of the l-th layer.

[0013] As a further aspect of the present invention, the Dropout regularization is implemented between the width expansion layer and the nonlinear transformation layer of the RKAN layer, and after the nonlinear transformation layer, with a dropout probability of 15%, to improve the generalization ability of the model.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By employing RKAN layers for deep feature transformation and regularization constraints, nonlinear correlation features in charging pile operation data are effectively extracted, enhancing the model's ability to represent complex health states. A bidirectional LSTM structure is used to simultaneously capture forward and backward temporal dependencies, fully understanding the evolution of health states over time and improving the early identification of fault symptoms. A multi-head self-attention mechanism is introduced to adaptively weight key time steps and features, enhancing the model's sensitivity to important health influencing factors and avoiding information overload. Finally, a health level probability distribution is output, facilitating maintenance personnel's understanding of the charging pile's health status and supporting preventative maintenance decisions. Attached Figure Description

[0015] Figure 1 This is a flowchart of a method for predicting the health of electric vehicle charging stations based on RKAN-BiLSTM-SA.

[0016] Figure 2 The diagram shows the RKAN-BiLSTM-SA network structure in the RKAN-BiLSTM-SA-based method for predicting the health of electric vehicle charging stations.

[0017] Figure 3 This is a diagram of the overall architecture of RKAN-BiLSTM-SA in the electric vehicle charging pile health prediction method.

[0018] Figure 4 The diagram shows the structure of the RKAN-BiLSTM-SA model in the RKAN-BiLSTM-SA-based method for predicting the health of electric vehicle charging stations. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0021] like Figure 1 and Figure 3 As shown, this embodiment of the invention provides a method for predicting the health of electric vehicle charging stations based on RKAN-BiLSTM-SA, the method comprising the following steps: S100, receive preprocessed time series input data, the input data includes historical health level data and seven key indicator features, the seven key indicators include load rate, actual charging power, cumulative charging amount, failure rate, time to use, average duration of failure and charging revenue; S200 inputs the input data into the RKAN layer for feature enhancement, including sequential width expansion, non-linear transformation and Dropout regularization. S300 inputs the features output from the RKAN layer into the BiLSTM-SA layer, extracts temporal features through a bidirectional long short-term memory network, and then uses a self-attention mechanism to weight and fuse the temporal features. S400 inputs the features output from the BiLSTM-SA layer into a fully connected layer for nonlinear transformation and dimensionality reduction. S500 outputs the prediction results of the future health of the charging pile through the output layer. The results are the probability distribution of the health level for multiple future time steps.

[0022] like Figure 2 As shown, it should be noted that the main technology enabling the implementation of this invention is the construction of an electric vehicle charging pile health prediction model based on RKAN-BiLSTM-SA. The RKAN-BiLSTM-SA model mainly consists of two parts: RKAN and BiLSTM-SA. The RKAN-BiLSTM-SA model adopts a hierarchical and progressive architecture, completing the charging pile health prediction task through the collaborative cooperation of four core modules. This architecture first performs deep processing on the input 14-dimensional time-series features, and finally outputs the probability distribution of five health levels.

[0023] In this embodiment of the invention, during the feature processing stage, the KAN feature transformation module performs a nonlinear transformation on the original features using a three-layer network structure. The RKAN layer comprises a first KANLayer, a Dropout layer, a second KANLayer, and a third KANLayer connected sequentially. The first KANLayer expands the feature dimension to 256 units, enhances feature representation with the ReLU activation function, and uses L2 regularization to control model complexity. After the Dropout layer randomly masks 15% of the neurons, the second KANLayer compresses the feature dimension to 128 units, further extracting deeper features. Finally, the third KANLayer unifies the output dimension, ensuring smooth integration with subsequent modules.

[0024] In this embodiment of the invention, the temporal feature extraction module employs a bidirectional long short-term memory network, including a forward LSTM and a backward LSTM, which simultaneously captures forward and backward temporal dependencies through 128 hidden units. This module retains the complete time-step output, providing fine-grained sequence features for subsequent attention mechanisms. During training, L2 regularization and Dropout techniques work together to effectively improve the model's generalization ability.

[0025] In this embodiment of the invention, the self-attention mechanism module first performs layer normalization on the temporal features, and then extracts key information from different feature subspaces through a 4-head self-attention mechanism; global average pooling converts the variable-length sequence into a feature vector of fixed dimensions; the fully connected layer further compresses the feature dimension to 64 units, and completes the nonlinear transformation through the ReLU activation function.

[0026] In this embodiment of the invention, the output layer receives the feature vector passed from the previous stage and calculates the probability of each health level using five output units and a Softmax function. Throughout the processing, each module performs its specific function: the feature transformation module mines deeper information from the original data, the time-series extraction module captures the evolutionary patterns of the health status, and the attention mechanism focuses on key time nodes, ultimately achieving an accurate assessment of the charging pile's health status.

[0027] In this embodiment of the invention, the overall architecture of RKAN-BiLSTM-SA consists of five layers: an input layer, an RKAN layer, a BiLSTM-SA layer, a fully connected layer, and an output layer.

[0028] The input layer is responsible for receiving data from the preprocessed data. All data is in time series format, containing T consecutive time steps. Specifically, the data received by this layer includes historical health level data that directly reflects the health status of the equipment. The system also includes multi-dimensional indicators related to the health of charging piles, covering three main categories of characteristics: operational efficiency, reliability, and economy. These include seven key indicators: load rate, actual charging power, cumulative charging volume, failure rate, time in use, average duration of failure, and charging revenue. The seven indicators of the time step constitute the feature vector. The overall formation of the time series feature matrix .

[0029] The RKAN layer receives data from the input layer and performs feature enhancement through the KAN network. The historical health status and three-level indicators of the charging pile constitute the input features. , The first layer undergoes width expansion through a linear transformation, the specific calculation of which is shown in the equation: ,in, This is the weight matrix. As a paranoia vector, The input features are projected into a higher-dimensional space to separate complex patterns; the second layer uses a nonlinear transformation, the specific calculation of which is shown in the equation: ,in, It is a non-linear activation function. , ; to achieve feature transformation and extraction.

[0030] Dropout regularization is applied between the width expansion layer and the nonlinear transformation layer, and after the nonlinear transformation layer. During model training, a certain percentage of neurons are randomly dropped with a dropout probability of 15%, which is used to improve the model's generalization ability. The specific operation of Dropout regularization is shown in the formula: , in, For the first Layer input; This represents the probability of dropping out (i.e., the proportion of neurons that are dropped out). This is the output after the Dropout regularization operation. By randomly discarding neuron connections, the model is forced to learn more robust feature representations, avoiding over-reliance on specific neurons.

[0031] The BiLSTM-SA layer, as the core temporal processing module of the model, plays a crucial role in connecting the preceding feature enhancement layer and the subsequent classification layer. This layer, through the collaborative work of BiLSTM and the SA mechanism, achieves deep mining of temporal features. Specifically, historical health scores, three-level indicators, and sentiment data are used as... The input is fed into a BiLSTM, which performs bidirectional processing on the input sequence. The forward LSTM captures the dependency between features and historical context, while the backward LSTM learns the association between features and future context. By concatenating the bidirectional hidden states, the features at each time step contain complete contextual information. The specific calculation is shown in the following equation: ; ; .

[0032] Building upon BiLSTM, the SA mechanism models global dependencies in sequences using a Query-Key-Value model. Three matrices are generated through linear transformation (learnable parameters are...). , (where the dimension is after transformation), the specific calculation is shown in the formula: ,in, These correspond to the query matrix, key matrix, and value matrix, respectively. The attention weight is obtained by calculating the similarity between the query vector and all key vectors, as shown in the formula below: ,in, yes The List, yes The Columns; organize all scores into a matrix ,in The attention scores are normalized using Softmax to obtain the attention weights. The specific calculation is shown in the formula: .

[0033] Then, the value vectors are weighted and fused to obtain the output for each position. This allows the new features at each time step to dynamically focus on the most relevant parts of the sequence, as shown in the formula: ,in, yes The This hierarchical processing mechanism enables the model to utilize both the local context modeling capabilities of BiLSTM and the global feature association established through the SA mechanism, thereby improving the recognition accuracy of complex temporal patterns.

[0034] The fully connected layer performs nonlinear transformation and dimensionality reduction on the high-dimensional temporal features extracted by BiLSTM-SA through the ReLU activation function, compressing the high-dimensional features output by the BiLSTM-SA layer to 64 dimensions, removing redundant information and retaining the most effective classification features.

[0035] The output layer receives the data after regularization by KAN-BiLSTM-SA and outputs the prediction result of the future health of the charging pile. The result is a time-series vector. ,in Each element in the vector represents the number of future time steps to be predicted. To predict the health level for the next future time step, where .

[0036] To balance the model's expressive power and generalization ability, an L2 regularization term is added to the loss function during the training of the RKAN-BiLSTM-SA model to implement constraints. The total loss function of the model is the sum of the task loss and the L2 regularization term, as shown in the formula: ,in, Cross-entropy loss for classification tasks ( For real labels, (for predicting probabilities) The regularization coefficient is used. This represents the number of layers in the model that contain the weight matrix. Indicates the first Layer weight matrix; For the first The squared L2 norm of the layer weight matrix.

[0037] like Figure 4 As shown in the embodiment of the present invention, the RKAN-BiLSTM-SA module structure diagram is mainly divided into three modules: the RKAN network realizes nonlinear expansion and regularization constraints of features through a multi-layer structure; the BiLSTM-SA network focuses on capturing temporal dependencies and strengthening key features; and the fully connected network finally outputs the probability distribution of health level.

[0038] In this embodiment of the invention, the RKAN network mainly consists of four modules: KAN_WE_Layer (width expansion) enhances feature representation capabilities through width expansion; KAN_NLT_Layer (nonlinear transformation) performs further nonlinear transformation on the features, refining key information strongly correlated with health status; ReLU, as the activation function, introduces nonlinearity, improving the network's ability to fit complex features; and Dropout regularization randomly discards some neurons, collectively enhancing the model's generalization ability.

[0039] In this embodiment of the invention, BiLSTM-SA is mainly divided into three modules. The BiLSTM network bidirectionally captures the temporal dependence of the charging pile's health status, considering both the impact of historical states on the current state and the potential correlation of future states. LayerNormalization normalizes the BiLSTM output, stabilizing the feature distribution and accelerating model training convergence. MultiHeadAttention uses a multi-head self-attention mechanism to capture key temporal features in multiple semantic spaces in parallel, adaptively amplifying the feature weights that play a decisive role in health prediction. GlobalAveragePooling1D aggregates sequence features into fixed-length vectors, compressing dimensionality while retaining global temporal information.

[0040] In this embodiment of the invention, the fully connected network is mainly divided into two modules. Dense performs linear transformation and nonlinear activation on the aggregated features to further refine the health status features. Softmax maps the features to a probability distribution of five health levels and outputs the final charging pile health prediction result.

[0041] The specific steps of the self-attention mechanism include: performing layer normalization on the feature sequence output by BiLSTM; generating a query matrix, key matrix, and value matrix through linear transformation; calculating attention weights and performing Softmax normalization; and performing weighted summation on the value matrix to obtain the attention output at each time step.

[0042] In this embodiment of the invention, the advantage of using the RKAN network in the implementation process is that it can perform deep feature extraction and nonlinear transformation on the input features, explore the complex correlation between indicators, and the combined use of L2 regularization and Dropout regularization can effectively prevent overfitting and improve the model's generalization ability. The advantage of using BiLSTM-SA is that BiLSTM can simultaneously capture the forward and reverse temporal dependencies of sequence data, fully explore the temporal correlation of features, understand the changing patterns of indicators over time, and the self-attention layer (SA) can assign different weights to the features output by BiLSTM, focus on key features, enhance attention to important information, and improve the model's ability to identify core factors.

[0043] In this embodiment of the invention, the advantages of using KAN_WE_Layer (width expansion) are: broadening the feature dimension and increasing the model's capacity to accommodate multiple features; the advantages of using KAN_NLT_Layer (nonlinear transformation) are: introducing nonlinearity, allowing the model to fit complex feature relationships; the advantages of using ReLU are: enhancing the model's nonlinear expressive power while solving the gradient vanishing problem of traditional activation functions, thus accelerating training; the advantages of using L2+Dropout regularization are: L2 regularization constrains the parameter size, and Dropout regularization randomly discards neurons; the combination of the two prevents overfitting from both the parameter and structural levels; and the use of LayerNormalization... The benefits of using tion are: normalizing features within a layer, accelerating model convergence, and improving training stability; using MultiHeadAttention is: capturing the correlation between features from multiple perspectives, more comprehensively mining key information, and enhancing the model's ability to focus on important features; using GlobalAveragePool1D is: performing global average pooling on sequence features, compressing feature dimensions while retaining global information, reducing the burden on subsequent fully connected layers; using Dense is: integrating features from all previous modules to prepare for the final output; and using Softmax is: transforming the output into a probability distribution, facilitating the interpretation of results in multi-class scenarios.

[0044] The above description only details the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0045] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0046] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A method for predicting the health of electric vehicle charging stations based on RKAN-BiLSTM-SA, characterized in that, The method includes the following steps: Receive preprocessed time series input data, which includes historical health level data and seven key indicator features, including load rate, actual charging power, cumulative charging amount, failure rate, time to use, average duration of failure, and charging revenue. The input data is fed into the RKAN layer for feature enhancement, including sequential width expansion, non-linear transformation, and Dropout regularization. The features output from the RKAN layer are input into the BiLSTM-SA layer, and temporal features are extracted through a bidirectional long short-term memory network. The temporal features are then weighted and fused through a self-attention mechanism. The features output from the BiLSTM-SA layer are input into a fully connected layer for nonlinear transformation and dimensionality reduction. The output layer outputs the prediction results of the future health of the charging pile, and the results are the probability distribution of the health level at multiple future time steps.

2. The method for predicting the health of electric vehicle charging piles based on RKAN-BiLSTM-SA according to claim 1, characterized in that, The RKAN layer comprises a first KANLayer, a Dropout layer, a second KANLayer, and a third KANLayer connected in sequence; wherein, the first KANLayer expands the input feature dimension to 256 dimensions and uses the ReLU activation function; the second KANLayer compresses the feature dimension to 128 dimensions; and the third KANLayer outputs a feature vector of uniform dimension for use by subsequent BiLSTM-SA layers.

3. The method for predicting the health of electric vehicle charging piles based on RKAN-BiLSTM-SA according to claim 1, characterized in that, The bidirectional long short-term memory network in the BiLSTM-SA layer includes a forward LSTM and a backward LSTM, which capture forward and backward temporal dependencies, respectively; the bidirectional LSTM has 128 hidden units and outputs the hidden state at each time step; the self-attention mechanism adopts a multi-head attention structure, including 4 attention heads, for extracting key information from different feature subspaces.

4. The method for predicting the health of electric vehicle charging piles based on RKAN-BiLSTM-SA according to claim 1, characterized in that, The specific steps of the self-attention mechanism include: Perform layer normalization on the feature sequence output by BiLSTM; Generate the query matrix, key matrix, and value matrix through linear transformation; Calculate the attention weights and perform Softmax normalization; The attention output for each time step is obtained by weighted summation of the value matrix.

5. The method for predicting the health of electric vehicle charging piles based on RKAN-BiLSTM-SA according to claim 1, characterized in that, The fully connected layer uses the ReLU activation function to compress the high-dimensional features output by the BiLSTM-SA layer to 64 dimensions and performs a nonlinear transformation.

6. The method for predicting the health of electric vehicle charging piles based on RKAN-BiLSTM-SA according to claim 1, characterized in that, The output layer uses the Softmax function to map the features output by the fully connected layer into a probability distribution of 5 health levels, representing the health status of the charging pile at each future time step.

7. The method for predicting the health of electric vehicle charging piles based on RKAN-BiLSTM-SA according to claim 1, characterized in that, During the training of the RKAN-BiLSTM-SA model, the total loss function consists of cross-entropy loss and L2 regularization term, specifically expressed as follows: , where L task The cross-entropy loss is λ, where λ is the regularization coefficient and W is the cross-entropy loss. l Let be the weight matrix of the l-th layer.

8. The method for predicting the health of electric vehicle charging piles based on RKAN-BiLSTM-SA according to claim 1, characterized in that, The Dropout regularization is implemented between the width expansion layer and the nonlinear transformation layer of the RKAN layer, and after the nonlinear transformation layer, with a dropout probability of 15%, to improve the generalization ability of the model.

Images