An electric bus charging station charging system anomaly prediction method

Abstract

This invention relates to the field of electronic digital data processing technology and discloses a method for predicting anomalies in electric bus charging station charging systems. The method includes: acquiring operational data, equipment data, and environmental monitoring data of the charging device; preprocessing the multi-source data to construct a high-dimensional time-series feature matrix; constructing and training a hierarchical ensemble learning model that integrates confidence perception, where the base learners include a time-series prediction model and a tree model, outputting prediction results and confidence scores; and a meta-learner performs weighted fusion based on the outputs of each base learner to output prediction results. The method further involves dividing scenarios according to region type and charging mode, training differentiated hierarchical ensemble learning models for each scenario to form a scenario-based model library; calling the corresponding model according to the target scenario to generate early warning information; and collecting newly labeled samples to incrementally learn the meta-learner, updating the model fusion weights and confidence parameters. This method solves the problems of slow anomaly identification and reliance on manual intervention, achieving the goals of early warning, high accuracy, and low cost.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a method for predicting anomalies in the charging system of an electric bus charging station. Background Technology

[0002] In the urban public transportation sector, pure electric buses have become the mainstream, and the scale of their supporting charging infrastructure is rapidly expanding. The charging system data acquisition system serving these bus charging stations / depots, as a key "nerve ending" of the smart grid, undertakes the core task of collecting crucial information such as power, voltage, current, and load from charging piles and transmitting it to the main station system. This data directly supports core business operations for bus companies, including accurate electricity billing, optimized charging scheduling, vehicle range assessment, and grid interaction. However, in actual operation, bus charging station systems have revealed increasingly prominent shortcomings in intelligent operation and maintenance. On the one hand, due to the complex environment of bus depots, the extremely high frequency of charging pile use, and the harsh operating conditions, problems such as unstable communication links, terminal disconnections, data loss, and data drift occur frequently, seriously affecting the integrity and timeliness of charging metering data. On the other hand, existing charging systems still rely heavily on manual inspections and experience-based judgment. Faced with a large number of charging piles and increasingly complex abnormal behaviors, it is difficult to detect problems in a timely manner and accurately locate faults. Once metering anomalies occur, they will not only lead to deviations in the electricity billing of bus companies, but may also affect vehicle charging plans and operation scheduling, falling into an inefficient closed loop of "delayed fault detection - difficulty in locating - slow response", becoming a bottleneck restricting the high-quality development of bus electrification.

[0003] For example, Chinese patent CN119961565A discloses a method and device for predicting the operating status of public charging stations, providing the following technical solution: a method and device for predicting the operating status of public charging stations, which to some extent solves the problem that most existing studies only consider charging and discharging data during charging station operation, failing to comprehensively consider multiple factors, and often use LSTM models alone for prediction, resulting in inaccurate prediction results. Furthermore, existing studies do not separate users' fast charging and slow charging needs, failing to fully meet user needs. The method for predicting the operating status of public charging stations includes: collecting training data; preprocessing the training data; training an initial model based on hyperparameter search and a source domain model; training a prediction model based on a target domain model; collecting prediction data and preprocessing it; and based on the prediction model and the preprocessed prediction data, obtaining predictions of available charging piles and occupancy rates. However, the above-mentioned method and device for predicting the operating status of public charging stations focuses on predicting charging pile availability and occupancy rates, without addressing the prediction problem of data collection anomalies, and lacks design for static equipment attributes, confidence perception, scene-differentiated modeling, and online incremental learning mechanisms. Summary of the Invention

[0004] This invention solves the problems of slow data anomaly identification, reliance on manual investigation, and insufficient model adaptability in the prior art. It proposes an anomaly prediction method for electric bus charging station charging systems, achieving the goals of early warning, high prediction accuracy, and low operation and maintenance costs.

[0005] Furthermore, this invention aims to predict and provide tiered early warnings for data collection anomalies in pure electric bus charging piles by integrating multi-source heterogeneous data and constructing a confidence-aware Stacking ensemble learning model, thereby ensuring the integrity and reliability of charging metering data and improving the stability of system operation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting anomalies in the charging system of an electric bus charging station includes: Acquire operational data, equipment data, and environmental monitoring data of the charging device, and preprocess the multi-source data to construct a high-dimensional time-series feature matrix; Construct and train a hierarchical ensemble learning model that integrates confidence awareness. The base learners include a temporal prediction model and a tree model, which output prediction results and confidence scores. The meta-learner performs weighted fusion based on the outputs of each base learner and outputs prediction results. Scenarios are divided according to region type and charging mode, and differentiated hierarchical ensemble learning models are trained for each scenario to form a scenario-based model library; The corresponding model is invoked based on the target scenario to generate early warning information; newly labeled samples are collected to incrementally learn the meta-learner and update the model fusion weights and confidence parameters.

[0007] By integrating multi-source data and utilizing a confidence-aware integrated model for scenario-based prediction and proactive early warning, the accuracy and adaptability of predicting charging system anomalies can be systematically improved, enabling intelligent operation and maintenance.

[0008] Preferably, the preprocessing of multi-source data includes: collecting time-series operation data of the charging device and completing the time axis; inserting empty records and marking missing data points; encoding and extracting features from the static attribute data of the equipment; mapping categorical variables to historical anomaly rates using target encoding; converting numerical attributes into runtime or production time difference as aging or batch risk features; and performing time alignment and normalization processing on environmental monitoring data.

[0009] This ensured the integrity and consistency of time-series data, avoided misjudgments by explicitly marking missing values, transformed static equipment attributes into features directly related to failure risks, and synchronized environmental data with equipment data, providing input data for subsequent modeling.

[0010] Preferably, the construction of the high-dimensional time series feature matrix includes aligning the time series operation data, equipment static attribute data, and environmental monitoring data according to the equipment and timestamp, extracting time series features including statistical features and rate of change features at a preset time scale, fusing them with static attributes and environmental features, and constructing a high-dimensional time series feature matrix as model input.

[0011] By aligning multi-source data and extracting time-series features, various data from the device are integrated, which can comprehensively reflect the overall operating status of the charging device during the time period and provide input data for subsequent modeling.

[0012] Preferably, in the hierarchical ensemble learning model that integrates confidence perception, the time-series prediction model includes an LSTM model and an Informer model, which use quantile regression to output predicted values ​​and their prediction variances; the tree model is an XGBoost model, which outputs predicted values ​​and uses the standard deviation of the decision tree ensemble as a confidence estimate; the confidence score of each base learner is calculated based on the prediction variance or standard deviation of its output, and is weighted and calibrated using historical prediction accuracy to obtain the final confidence score.

[0013] An uncertainty estimation mechanism was designed for different types of base learners, and the confidence level was dynamically calibrated in combination with the historical performance of the model, so that the system can better evaluate the reliability of each prediction.

[0014] Preferably, the meta-learner adopts a linear regression model with L2 regularization. The input is a meta-feature vector formed by concatenating the predicted values ​​of each base learner and their final confidence scores. The fusion weights of each base learner are learned through training, and a confidence-based weighted fusion strategy is adopted to give the output of the high-confidence model a higher weight in the final prediction.

[0015] By automatically learning how to dynamically weight the prediction results of each base learner based on their reliability, the meta-learner integrates the advantages of different models, suppresses the negative impact of low-quality predictions, and thus improves the stability and accuracy of the overall prediction model.

[0016] Preferably, the scenario division based on region type and charging mode includes dividing the training dataset into multiple subsets according to the geographical and climatic characteristics of the device deployment area and the charging mode. Each subset corresponds to a type of operating scenario. A dedicated hierarchical ensemble learning model is independently trained for each type of scenario. The parameters, feature encoding mapping tables, and confidence calibration coefficients of each model are stored separately to form a scenario-based model library. During the prediction stage, the corresponding model is automatically called for inference based on the region and charging mode label of the target device.

[0017] By establishing specialized models for different geographical climates and charging modes, the problem of insufficient generalization ability of general models in diverse real-world scenarios is solved. These models can better fit data patterns and anomalies in specific scenarios, significantly improving the relevance and accuracy of predictions.

[0018] Preferably, the generation of early warning information includes classifying early warning into three levels based on the comprehensive prediction missing time and average confidence level output by the model; the early warning information is pushed to the operation and maintenance platform through a message protocol, and the early warning information includes the device identifier, the predicted interruption duration and confidence level, and the suggested processing time.

[0019] Preferably, the incremental learning includes periodically collecting manually verified samples to form a new labeled dataset, adjusting the fusion weights of the meta-learners based on the new labeled dataset, updating the historical prediction accuracy of each base learner, and recalibrating the weighting coefficients and decay coefficients in the confidence score calculation.

[0020] The model can continuously optimize using new data generated in actual operation and maintenance, dynamically adjust the model fusion strategy and confidence evaluation parameters, and enable the system to adapt to long-term evolution.

[0021] Preferably, the incremental learning also includes a model parameter rollback mechanism, which retains a backup of the old version after each model update. If the prediction accuracy of the new model on the validation set drops below a set threshold, it automatically rolls back to the previous stable version.

[0022] The rollback mechanism provides security for online model updates, ensuring that even if the performance of the new model degrades due to data fluctuations or other reasons during incremental learning, the system can automatically revert to a stable version, guaranteeing the continuity and reliability of long-term operation in the production environment.

[0023] Preferably, both the time-series prediction model and the tree model use a high-dimensional time-series feature matrix as input; the LSTM model adopts a two-layer stacked structure, with two neurons in the output layer that output the predicted missing time and its variance respectively, and the loss function is negative log-likelihood; the Informer model includes an encoder and a decoder, and uses the ProbSparse self-attention mechanism to handle long-period dependencies, with an output structure consistent with LSTM; the XGBoost model outputs a single predicted value and generates a confidence estimate based on the standard deviation of the prediction results of all decision trees for the same input sample.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0025] 1. This invention constructs a multi-dimensional feature system by integrating the time-series operation data of the charging device, the static attributes of the device, and environmental monitoring information. It adopts an ensemble learning framework based on confidence perception, which enables the model to dynamically adjust the fusion weights according to the prediction confidence of each base learner, giving priority to high-confidence outputs and effectively suppressing low-quality prediction interference, thereby significantly improving prediction accuracy and overall robustness.

[0026] 2. This invention uses the missing time prediction value output by the model to determine anomalies and automatically generates graded early warning information to be pushed to the operation and maintenance platform in real time. This guides personnel to intervene in potential faults in advance, which not only greatly improves the timeliness of fault detection and handling, but also significantly reduces the cost and lag of relying on manual inspections, thus ensuring the stable operation of the charging system.

[0027] 3. This invention establishes differentiated prediction models based on region type and charging mode, which fully adapts to the data characteristics and anomaly patterns under different power consumption scenarios. At the same time, the system supports online incremental learning using newly labeled data on a regular basis, dynamically adjusting model parameters and fusion weights, so that the model can continuously adapt to factors such as equipment aging and environmental changes, and maintain the long-term effectiveness and stability of prediction performance. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of the method for predicting anomalies in the charging system of an electric bus charging station according to the present invention.

[0029] Figure 2 This is a schematic diagram of the multi-source heterogeneous data fusion and feature engineering structure of the electric bus charging station charging system anomaly prediction method of the present invention. Detailed Implementation

[0030] See Figures 1 to 2 As shown, a method for predicting anomalies in an electric bus charging station system includes: Acquire operational data, equipment data, and environmental monitoring data of the charging device, and preprocess the multi-source data to construct a high-dimensional time-series feature matrix; Construct and train a hierarchical ensemble learning model that integrates confidence awareness. The base learners include a temporal prediction model and a tree model, which output prediction results and confidence scores. The meta-learner performs weighted fusion based on the outputs of each base learner and outputs prediction results. Scenarios are divided according to region type and charging mode, and differentiated hierarchical ensemble learning models are trained for each scenario to form a scenario-based model library; The corresponding model is invoked based on the target scenario to generate early warning information; newly labeled samples are collected to incrementally learn the meta-learner and update the model fusion weights and confidence parameters.

[0031] To address the issues of delayed data acquisition anomaly identification and reliance on manual investigation in existing charging systems serving pure electric bus charging stations, this invention aims to provide a method for predicting data acquisition anomalies. This method integrates multi-source heterogeneous data and constructs an ensemble learning prediction model to achieve proactive prediction and accurate early warning of charging pile data acquisition anomalies. This improves the automation and intelligence of anomaly identification, ensures the accuracy and reliability of bus charging metering data, reduces operation and maintenance costs, and provides reliable data support for the refined management and safe, stable operation of the public transportation system.

[0032] like Figure 1 In one embodiment shown, Figure 1 This is an overall flowchart of an anomaly prediction method for an electric bus charging station charging system according to the present invention. First, the operating data of the charging device, equipment static attributes, and environmental monitoring data are acquired and preprocessed to construct a high-dimensional time-series feature matrix as model input. The equipment static attributes are mapped to historical anomaly rates using target encoding, and the installation date is converted to runtime as an aging feature. Environmental monitoring data is aligned and then normalized using Min-Max. Statistical features and rate of change features are extracted by aligning the equipment with the timestamp, and the static and environmental data are then fused.

[0033] The model employs a hierarchical ensemble learning framework that incorporates confidence awareness. The base learners include the temporal prediction models LSTM and Informer, as well as the tree model XGBoost. All of them take a high-dimensional temporal feature matrix as input. LSTM uses a two-layer stacked structure to output the predicted missing time and its variance, and uses a negative log-likelihood loss function. Informer includes an encoder and a decoder and uses a ProbSparse self-attention mechanism to handle long-term dependencies. Both output the predicted value and variance through quantile regression. XGBoost outputs a single predicted value and uses the standard deviation of the decision tree ensemble as the confidence estimate. The confidence score of each base learner is calculated based on the variance or standard deviation of the output and is calibrated by weighting with historical prediction accuracy to obtain the final confidence score.

[0034] The meta-learner is a linear regression model with L2 regularization. It takes as input a meta-feature vector concatenated from the predictions of each base learner and the final confidence score. Through training, it learns fusion weights and employs a confidence-based weighted fusion strategy to give higher-confidence model outputs greater weight. To improve adaptability, scenarios are segmented based on region type and charging mode. The training dataset is divided into subsets based on the geographical and climatic characteristics of the device deployment area and the charging mode, and dedicated hierarchical ensemble learning models are trained independently to form a scenario-based model library. During prediction, the appropriate model is automatically invoked for inference based on the target device's region and charging mode label.

[0035] The model outputs a comprehensive prediction of missing time and average confidence level, based on which a three-level early warning system is established, generating early warning information including device identifier, prediction interruption duration, confidence level, and suggested processing time, which is pushed to the operation and maintenance platform via a message protocol. Furthermore, this invention periodically performs incremental learning, collecting manually confirmed abnormal samples to form a new labeled dataset. Based on this, the meta-learner fusion weights are adjusted, the historical prediction accuracy of each base learner is updated, and confidence parameters such as weighting coefficients and decay coefficients are recalibrated. Simultaneously, a model parameter rollback mechanism is introduced, retaining a backup of the old version after each update. If the prediction accuracy of the new model on the validation set drops beyond a set threshold, it automatically rolls back to the previous stable version.

[0036] This invention provides a method for anomaly prediction of a pure electric bus charging station charging system based on multi-source data fusion and scenario adaptation, comprising the following steps: Step 1, Multi-source heterogeneous data fusion and feature engineering: Simultaneously collect time-series data such as voltage, current, power factor, and positive active power of the charging device for pure electric buses, static attribute data such as equipment asset number, manufacturer, production batch, and installation date, as well as monitoring data such as temperature and humidity of the transformer area; perform missing value filling, outlier cleaning, time alignment and normalization on the data, and construct a high-dimensional time-series feature matrix through feature cross and time-series sliding window techniques.

[0037] Step 2, Building the Stacking Ensemble Learning Model with Confidence Scores: The base learner layer includes an LSTM (Long Short-Term Memory) model, an Informer (efficient long-sequence time series prediction) model, and an XGBoost model. The meta-learner is a linear regression model with L2 regularization. The Stacking ensemble learning model with confidence scores is then constructed.

[0038] Furthermore, the improvement of the Stacking ensemble learning model that integrates confidence scores is mainly divided into the following three aspects; Step A: Introduce an uncertainty output mechanism into the LSTM and Informer models so that the models not only output the predicted missing time, but also reflect the credibility of their prediction results, providing a reliable basis for subsequent integrated decision-making and avoiding the misleading effect of high-error predictions on the final results.

[0039] Step B: Generate dynamic confidence scores for each base learner to comprehensively evaluate the predictive reliability of each model under the current input conditions. This solves the problem of unreasonable fusion caused by "equal weighting" or "fixed weighting" of all models in traditional ensemble methods, and improves the adaptability of the ensemble process.

[0040] Step C: At the meta-learner layer, a confidence-based weighted fusion strategy is adopted. The contribution weight of each base learner in the final prediction is automatically adjusted according to the confidence of each base learner. The output of the high-confidence model is adopted first, and the interference of low-reliability prediction is suppressed, thereby improving the stability and prediction accuracy of the overall model.

[0041] Step 3, Scene Adaptive Model Training: Divide the dataset according to region type and charging mode, and train a dedicated Stacking model independently for each scene type to realize scene-specific configuration of model parameters and feature weights, thereby improving prediction accuracy.

[0042] Step 4, Dynamic Early Warning Deployment and Online Incremental Learning: The trained differentiated model is deployed in the real operating environment, and the data stream of the pure electric bus charging system is accessed in real time. Based on the real operating environment, the differentiated ensemble model from Step 3 is executed. Anomaly prediction results are output through feature extraction and model inference, thereby generating early warning information and pushing it to the operation and maintenance platform to guide maintenance personnel to intervene in advance. The meta-learner layer is periodically updated incrementally using newly collected labeled data, enabling the model to continuously learn new features and anomaly patterns in system operation, dynamically adjust the fusion weights of each base learner, adapt to long-term evolutionary factors such as equipment aging and environmental changes, and maintain the stability and long-term effectiveness of predictive performance.

[0043] This invention provides a method for predicting data acquisition anomalies in pure electric bus charging station charging systems based on multi-source data fusion and scenario adaptation. It constructs a multi-dimensional feature system by fusing time-series operational data of charging devices, static equipment attributes, and environmental monitoring information. A confidence-aware Stacking ensemble learning framework is employed, enabling the meta-learner to dynamically adjust fusion weights based on the prediction reliability of each base learner, prioritizing the output of high-confidence models, effectively suppressing interference from low-quality predictions, and significantly improving the accuracy and robustness of anomaly prediction. Furthermore, by establishing differentiated prediction models based on region type and charging mode, it fully adapts to data from different electricity consumption scenarios. The differences in features and anomaly patterns solve the problem of insufficient generalization ability caused by "one-size-fits-all" modeling. Anomaly detection is performed based on the missing time prediction values ​​output by the model, and hierarchical early warning information is generated and pushed to the operation and maintenance platform. This realizes the transformation of the operation and maintenance mode from "post-event response" to "pre-event early warning and precise intervention", which improves the efficiency of fault response. At the same time, by regularly using newly labeled data to incrementally update the model, the system can dynamically adapt to long-term evolution factors such as equipment aging and environmental changes, maintain the continuous stability of prediction performance, extend the model life cycle, significantly reduce the cost of manual inspection, and effectively improve the intelligent operation and maintenance level of the charging system of pure electric bus charging stations.

[0044] In another embodiment, the present invention provides a method for predicting data acquisition anomalies in a pure electric bus charging station charging system based on multi-source data fusion and scenario adaptation, comprising the following steps: Step 1, Multi-source data fusion preprocessing and temporal feature matrix construction, refer to Figure 2 , Figure 2 This diagram illustrates the multi-source heterogeneous data fusion and feature engineering structure of an anomaly prediction method for electric bus charging stations according to the present invention. The invention first collects, cleans, and fuses multi-source heterogeneous data from the charging system of a pure electric bus charging station to construct a structured, high-quality input feature set, providing a reliable data foundation for subsequent model training.

[0045] Step 1-1: Collection and Processing of Time-Series Operational Data for Charging Devices: This step involves collecting operational data from the charging devices of pure electric buses, including voltage, current, power factor, forward active power, and daily frozen power consumption. The standard reporting cycle is one record every 15 minutes. To ensure the continuity of the time-series data, the data stream of each charging device is first time-tracked. Specifically, a complete time series (i.e., timestamp set: t0, t0+15min, t0+30min, ...) is constructed for each device, from the start time to the end time with a step size of 15 minutes, and aligned with the actual received data records. For missing data points on this timeline (i.e., no corresponding reported record), the system automatically inserts an empty record and uniformly fills the original data fields such as voltage, current, and power with "-1" as a clear missing marker, avoiding misjudgment as valid zero or low values ​​in subsequent processing.

[0046] Steps 1-2: Encoding and Feature Extraction of Static Attribute Data for Equipment: Collect static attributes for each charging device, including asset number, manufacturer, production batch, installation date, equipment model, and communication module type. For categorical variables such as "manufacturer" and "equipment model," a target encoding method is used to map them to the historical anomaly rate of that category of equipment. For "installation date," calculate the equipment runtime T. age equal to t current Subtract t install (Unit: days), as a characteristic of equipment aging. For "production batches", it is converted to the production start time, and the difference between the start time and the current time is calculated to identify batch failure risks.

[0047] Steps 1-3: Alignment and Normalization of Environmental Monitoring Data: Collect temperature, humidity, and other environmental data for the charging device. Environmental data is typically sampled hourly or daily and needs to be aligned with the charging device data by timestamp using a nearest neighbor matching strategy. Numerical features (such as temperature and humidity) undergo Min-Max normalization, specifically: x' equals x minus x. min Difference divided by xmax Subtract x min The difference, where x is the original numerical feature value, x max x represents the maximum value of this feature over the time scale. min x' represents the minimum value of the feature over the time scale, and x' is the normalized feature value. This ensures that all features are on the same scale, avoiding numerical instability during model training.

[0048] Steps 1-4, Multi-source Data Fusion and Feature Engineering: Align the above three types of data according to "equipment asset number and timestamp" to form a unified data table. Based on this, perform time-series feature engineering: Use 24 hours as a time scale for time-series feature processing. If more than 50% of the data rows within a time scale are missing, assign a uniform value of "-1" to the features within that time scale; otherwise, ignore the missing data rows and extract statistical features (such as the average voltage V). 均 Maximum voltage V max Minimum voltage V min Standard deviation σ V σ V 2 Other dimensional features are also processed in the same way), and the rate of change features (such as the difference in the total positive active power change Δpap_r equal to pap_r) are also processed in the same way. t Subtract pap_r t-1 Then, by incorporating the feature data from steps 1-2 and 1-3, a high-dimensional time-series feature matrix Χ∈R is finally constructed. N×F , where N is the total number of samples and F is the feature dimension, which serves as the model input.

[0049] Step 2: Constructing a confidence-aware Stacking ensemble learning model. This invention constructs a two-layer Stacking ensemble learning model, which improves the accuracy and robustness of prediction by introducing confidence scores to achieve dynamic weighted fusion.

[0050] Step 2-1, Training and Uncertainty Output of the Base Learner Layer: The base learner layer consists of three modules: LSTM, Informer, and XGBoost. The LSTM module is designed to capture short-to-medium-term temporal features and perform uncertainty estimation. This module adopts a two-layer stacked structure, with each layer containing 256 hidden units, and uses Tanh as the activation function to maintain gradient stability. After the LSTM feature extraction layer, a fully connected layer containing 256 neurons is connected, using the ReLU activation function to increase non-linear mapping capability. During training, the initial learning rate is set to 0.001, the batch size is 64, the maximum training epochs are 100, and an early stopping strategy is adopted, with a patience value set to 10. The output layer is configured with two neurons, outputting the predicted missing time and its prediction variance σ, respectively. 2This achieves quantile regression. Its loss function uses negative log-likelihood, specifically: L... LSTM equal to half of log(2πσ) 2 ), and in addition, (t true -t 持 ) 2 Divide by 2σ 2 Among them, L LSTM Let t be the loss value of the LSTM module. 持 t is used to predict the duration of data acquisition failures for the module. true σ represents the actual duration of data acquisition failures. 2 This represents the prediction variance of the module.

[0051] The Informer module handles long-cycle dependencies. Its core architecture consists of an encoder and a decoder, both composed of three stacked Informer blocks. The sampling factor of the ProbSparse self-attention mechanism in each block is set to 5. The hidden dimension of the module is set to 512, the feedforward network dimension is set to 2048, the number of attention heads is set to 8, and the input and output are consistent with the LSTM module.

[0052] The XGBoost module takes the constructed high-dimensional feature matrix as input, and like the LSTM and Informer modules, it outputs a single predicted value t. XGB The standard deviation of the decision tree ensemble is used as a confidence estimate. The hyperparameters used in this paper are as follows: maximum tree depth is 8, learning rate is 0.1, subsample ratio is 0.8, column sampling ratio is 0.8, regularization parameters L1 is 0, L2 is 1, and the number of parallel threads used to build each tree is -1.

[0053] Step 2-2, Generation and calibration of dynamic confidence scores: Generate a confidence score for each base learner.

[0054] For LSTM and Informer, the confidence level is: C i It equals exp (-λ multiplied by σ'), where C i Let σ' be the initial confidence score of the i-th base learner, σ' be the prediction standard deviation of the model output, and λ be the decay coefficient with a value of 0.5.

[0055] For XGBoost, the confidence level is: C XGB It equals 1 / (1 plus σ) tree ), where C XGB σ is the initial confidence score for the XGBoost model. tree is the standard deviation of the prediction results of all decision trees for the same input sample.

[0056] The system maintains the historical prediction accuracy (Acc) of each model over a past period. i (defined as the proportion of prediction errors less than 20%), and the final confidence level is calibrated using a weighted method, specifically: C i final Equals α multiplied by C i In addition, (1-α) multiplied by Acc i , where C i final Acc is the final confidence score for the i-th base learner. i This represents the proportion of samples where the model's prediction error is less than 20% over a past period, with α being a weighting coefficient of 0.7. This mechanism ensures that the model automatically reduces its weights when its performance deteriorates.

[0057] Steps 2-3, Weighted Fusion and Training of Meta-Learners: The prediction results and final confidence scores of each base learner are concatenated to form a meta-feature vector, specifically: Χ meta equals [t] LSTM C LSTM final , t Informer C Informer final , t XGB C XGB final ].

[0058] A linear regression model with L2 regularization is used as the meta-learner to learn a weighted fusion strategy, specifically: y' equals ∑ i=1 3 ω i ∙(t i •C i final Where y' represents the duration of missing data collection for the final ensemble learning prediction, and t i Let C be the predicted value of the i-th base learner. i final ω is the final confidence score for the i-th base learner. i These are the trainable weights learned through meta-learning.

[0059] During training, k-fold cross-validation is used to generate meta-features, avoiding data leakage. This strategy ensures that the output of the high-confidence model dominates the final prediction, improving overall robustness.

[0060] Step 3, Scene Adaptive Model Training: To improve the model's adaptability to different operating environments, this invention adopts a scene adaptive modeling strategy to achieve fine-grained configuration of model parameters and feature weights.

[0061] Step 3-1: Divide the dataset according to regional geographical climate type and charging mode: Divide the training dataset according to the geographical climate characteristics of the device deployment area (such as the hot and humid areas in the south and the cold and dry areas in the north) and the charging mode. Construct a sub-dataset for each scenario to ensure the consistency of data distribution.

[0062] Step 3-2: Independently train scenario-specific prediction models: For each subset of datasets, independently execute the model building process in Step 2 to train a dedicated Stacking ensemble model. After training, the model parameters, feature encoding mapping tables, and confidence calibration coefficients are stored separately to form a scenario-specific model library.

[0063] Step 3-3: Model Selection and Invocation Mechanism: During the deployment phase, the system automatically invokes the corresponding dedicated model from the model library for inference based on the "region" and "charging mode" labels of the device to be predicted, ensuring that the prediction results are highly matched with the actual operating environment and improving the model's generalization ability and prediction accuracy.

[0064] Step 4, Dynamic Early Warning Deployment and Online Incremental Learning: This invention deploys the trained model in a real operating environment to achieve closed-loop management from prediction to intervention, and ensures the long-term effectiveness of the model through an online learning mechanism.

[0065] Step 4-1, Real-time Inference and Anomaly Detection: Deploy the trained differentiated model in a real business environment and receive real-time data streams from the pure electric bus charging system. The system performs data preprocessing and feature extraction as in Step 1, calls the model corresponding to the scenario for inference, and outputs a comprehensive prediction of the missing time.

[0066] Step 4-2, Generation and Push of Tiered Early Warning Information: Generate a three-level early warning based on the prediction results. If y' ≥ 24 hours and the overall confidence level C avg If the value is greater than 0.6, a "high-risk" warning is generated, and immediate action is recommended; if 20 < y' < 24 hours, a "medium-risk" warning is generated, and action is recommended within 24 hours; no warning is issued in other cases. Warning information is pushed to the operations and maintenance management platform via the MQTT protocol, including the device ID, predicted outage duration, confidence level, and recommended handling time, guiding operations and maintenance personnel to intervene precisely.

[0067] Step 4-3, Regular Incremental Learning and Model Update: The system is set to an incremental learning cycle every two weeks, automatically collecting abnormal samples confirmed by operations and maintenance personnel within the past 15 days to form a new labeled dataset. This data is used to adjust the fusion weights ω of the meta-learner. i Fine-tuning was performed using gradient descent with a small learning rate. Simultaneously, the historical accuracy (Acc) of each base learner was updated. i And recalibrate the confidence score parameters α and λ.

[0068] Step 4-4, Model Parameter Rollback Mechanism: After the update is complete, the system automatically switches to the new model version, while the old version is retained as a backup. If the performance of the new model on the validation set degrades beyond a threshold (e.g., accuracy drops by >3%), the automatic rollback mechanism is triggered to restore the previous stable version, ensuring the continuity and reliability of system operation. This mechanism enables the model to continuously adapt to long-term evolutionary factors such as equipment aging, network upgrades, and seasonal changes, significantly extending the model's lifespan.

[0069] All data collection and extraction in this invention are carried out under compliant and legal conditions.

Claims

1. A method for predicting anomalies in the charging system of an electric bus charging station, characterized in that, include: Acquire operational data, equipment data, and environmental monitoring data of the charging device, and preprocess the multi-source data to construct a high-dimensional time-series feature matrix; Construct and train a hierarchical ensemble learning model that integrates confidence awareness. The base learners include a temporal prediction model and a tree model, which output prediction results and confidence scores. The meta-learner performs weighted fusion based on the outputs of each base learner and outputs prediction results. Scenarios are divided according to region type and charging mode, and differentiated hierarchical ensemble learning models are trained for each scenario to form a scenario-based model library; The corresponding model is invoked based on the target scenario to generate early warning information; Collect newly labeled samples to incrementally learn the meta-learner and update the model fusion weights and confidence parameters.

2. The method for predicting anomalies in an electric bus charging station charging system according to claim 1, characterized in that, The preprocessing of multi-source data includes: collecting time-series operation data of charging devices, completing the time axis, inserting empty records and marking missing data points; encoding and extracting features from static attribute data of equipment, mapping categorical variables to historical anomaly rates using target encoding, and converting numerical attributes into runtime or production time difference as aging or batch risk features. Environmental monitoring data are time-aligned and normalized.

3. The method for predicting anomalies in an electric bus charging station charging system according to claim 1 or 2, characterized in that, The construction of the high-dimensional time series feature matrix includes aligning the time series operation data, equipment static attribute data, and environmental monitoring data according to the equipment and timestamp, extracting time series features including statistical features and rate of change features at a preset time scale, fusing them with static attributes and environmental features, and constructing a high-dimensional time series feature matrix as model input.

4. The method for predicting anomalies in an electric bus charging station charging system according to claim 3, characterized in that, In the hierarchical ensemble learning model that integrates confidence perception, the time-series prediction model includes an LSTM model and an Informer model, which use quantile regression to output predicted values ​​and their prediction variances; the tree model is an XGBoost model, which outputs predicted values ​​and uses the standard deviation of the decision tree ensemble as a confidence estimate; the confidence score of each base learner is calculated based on its output prediction variance or standard deviation, and is weighted and calibrated using historical prediction accuracy to obtain the final confidence score.

5. The method for predicting anomalies in an electric bus charging station charging system according to claim 4, characterized in that, The meta-learner employs a linear regression model with L2 regularization. The input is a meta-feature vector composed of the predicted values ​​of each base learner and their final confidence scores. The fusion weights of each base learner are learned through training, and a confidence-based weighted fusion strategy is adopted to give higher weight to the output of the high-confidence model in the final prediction.

6. The method for predicting anomalies in an electric bus charging station charging system according to claim 5, characterized in that, The scenario division based on region type and charging mode includes dividing the training dataset into multiple subsets according to the geographical and climatic characteristics of the device deployment area and the charging mode. Each subset corresponds to a type of operating scenario. A dedicated hierarchical ensemble learning model is independently trained for each type of scenario. The parameters, feature encoding mapping table and confidence calibration coefficient of each model are stored separately to form a scenario-based model library. During the prediction phase, the corresponding model is automatically invoked for inference based on the target device's region and charging mode label.

7. The method for predicting anomalies in an electric bus charging station charging system according to claim 6, characterized in that, The generation of early warning information includes classifying early warning into three levels based on the comprehensive prediction missing time and average confidence level output by the model; the early warning information is pushed to the operation and maintenance platform through a message protocol, and the early warning information includes the device identifier, the predicted interruption duration and confidence level, and the suggested processing time.

8. The method for predicting anomalies in an electric bus charging station charging system according to claim 7, characterized in that, The incremental learning includes periodically collecting manually verified samples to form a new labeled dataset, adjusting the fusion weights of the meta-learners based on the new labeled dataset, updating the historical prediction accuracy of each base learner, and recalibrating the weighting coefficients and decay coefficients in the confidence score calculation.

9. A method for predicting anomalies in an electric bus charging station charging system according to claim 7 or 8, characterized in that, The incremental learning also includes a model parameter rollback mechanism, which retains a backup of the old version after each model update. If the prediction accuracy of the new model on the validation set drops below a set threshold, it will automatically roll back to the previous stable version.

10. The method for predicting anomalies in an electric bus charging station charging system according to claim 4, characterized in that, Both the time-series prediction model and the tree model take a high-dimensional time-series feature matrix as input. The LSTM model adopts a two-layer stacked structure, with two neurons in the output layer that output the predicted missing time and its variance respectively. The loss function is negative log-likelihood. The Informer model includes an encoder and a decoder, and uses the ProbSparse self-attention mechanism to handle long-period dependencies. Its output structure is consistent with that of LSTM. The XGBoost model outputs a single predicted value and generates a confidence estimate based on the standard deviation of the prediction results of all decision trees for the same input sample.

Images