A charging station load prediction data processing method and system based on dynamic fusion of multi-source heterogeneous data

Abstract

The present application relates to a kind of based on the dynamic fusion of multi-source heterogeneous data charging station load prediction data processing method and system, belong to charging station electric power management technical field.The method is first by pluggable data source adapter mode access charging station historical power, real-time weather and charging pile attribute such as multi-source heterogeneous data;Then sequentially execute verification, alignment, cleaning and normalization processing to data;Through the heterogeneous expansion strategy of containing time window expansion, multi-site generation, noise injection enhancement, mode resampling, cross-season migration and weather-load joint sample, sample expansion and fusion are carried out, and the proportion of fusion data set and real data is adjusted based on reinforcement learning-based fusion proportion optimization;Finally, train time series prediction model on optimal fusion data set, and based on closed-loop feedback mechanism, continuously iteratively optimize.The present application can effectively solve the problem of training sample scarcity and multi-source data cooperation in charging station, significantly improve the load prediction accuracy in cold start scenario.

Description

Technical Field

[0001] This invention belongs to the field of charging station power management technology, and relates to a charging station load forecasting data processing method and system based on dynamic fusion of multi-source heterogeneous data. Background Technology

[0002] With the rapid growth of electric vehicle ownership and the large-scale application of high-power fast charging technology, the large-scale integration of charging stations into the power distribution network has become a typical scenario for the construction of new power systems. By the end of 2025, the number of electric vehicles in China had exceeded 30 million, and the scale of charging infrastructure exceeded 12 million units, including more than 300,000 public charging stations. As a key hub connecting electric vehicles and the power distribution network, the accurate prediction of the operating load of charging stations plays an important supporting role in the safe operation of the power distribution network, orderly charging scheduling, transformer capacity planning, and electricity market transactions. Accurate load forecasting can achieve the following: (a) providing power distribution network dispatching agencies with trends in charging load changes, assisting in the formulation of orderly charging strategies, and avoiding transformer overload and line congestion; (b) enabling charging station operators to optimize charging strategies under time-of-use pricing, reducing electricity purchase costs and increasing operating revenue; (c) providing a load baseline for V2G reverse power supply scheduling, supporting electric vehicles to participate in the grid ancillary services market; and (d) providing a basis for capacity planning for the construction and expansion of charging stations, avoiding over-investment or insufficient capacity.

[0003] However, charging station load forecasting faces a series of unique technical challenges that differ from traditional power load forecasting. These challenges stem from the combined effects of the charging station's operating characteristics, data environment, and external influencing factors, specifically including the following four aspects.

[0004] (i) Heterogeneity of data sources and complexity of access: Load forecasting of charging stations needs to integrate heterogeneous data from multiple sources, such as historical power data, real-time meteorological data (multiple API providers), charging pile attribute data and electricity price period data. The data sources differ significantly in terms of interface protocol, data structure, sampling frequency and data quality.

[0005] (ii) Severe scarcity of training samples: Most charging stations have been in operation for less than six months. With a sampling interval of 5 minutes, only about 25,920 data points are accumulated in 3 months. Deep learning models usually require more than 50,000 data points, which leads to problems such as overfitting and lack of extreme scenarios.

[0006] (III) Multi-scale composite characteristics of charging load: The load of charging stations exhibits composite characteristics across four time scales: impact, rhythm, cycle, and season. Traditional single data augmentation methods can only affect the impact scale and cannot provide effective extended coverage on the other three scales.

[0007] (iv) Lack of multi-strategy coordination and adaptive fusion: The effectiveness of different expansion strategies varies significantly in different scenarios. Existing technologies lack an adaptive optimization mechanism for multi-strategy coordination scheduling and fusion ratio.

[0008] To address the aforementioned problems, some research has been conducted in the existing technical field (such as CN116596129A, CN116629442A, CN110119841A, CN118645992A, CN118336678A, etc.), and their main technical solutions and shortcomings are analyzed as follows: Among them, CN116596129A (Shenzhen Power Supply Bureau Co., Ltd.) proposed an ALSTM-SE-CNN dual-branch prediction model, but data augmentation relies on a limited number of manually constructed auxiliary variables and does not involve a systematic process for unified access to multi-source heterogeneous data and multi-strategy collaborative augmentation. CN116629442A (Sun Yat-sen University) uses VMD decomposition + LSTM component-by-component prediction, but it only uses a single historical load data source, does not uniformly access multi-source heterogeneous data, and lacks a multi-strategy collaborative sample augmentation scheme.

[0009] CN110119841A (Guangdong University of Technology) proposes a simulation framework for user charging behavior based on a discrete choice model, but it does not involve unified access and standardized processing of multi-source data, and lacks an end-to-end learning mechanism and adaptive data augmentation scheme. CN118645992A (Hubei Complex Spacetime Energy Technology Co., Ltd., etc.) uses IACO to optimize LSTM hyperparameters, but the input variables only include electricity price and holidays, and it does not establish an integrated mechanism of "multi-strategy expansion - reinforcement learning optimization - closed-loop feedback". CN118336678A (State Grid Shanghai Electric Power Company) uses LASSO+BP neural network for medium-term load forecasting, but it does not provide a systematic multi-source heterogeneous data access framework, and the fusion process relies on manual feature engineering, making it difficult to adapt to the data distribution of different charging stations.

[0010] In summary, existing technologies address the following technical problems in current charging station load forecasting technologies: (1) Lack of a unified framework for multi-source heterogeneous data access: Existing technologies have different interfaces, tight coupling, and poor scalability when processing multi-source heterogeneous data such as historical power data, real-time meteorological data, and charging pile attribute data. The logic of establishing connections, authentication handshakes, data acquisition, and protocol conversion for each data source is hard-coded in the business modules. Adding or changing data sources requires modifying multiple core code entries, resulting in long development and testing cycles and high system maintenance costs.

[0011] (2) Training samples are severely scarce and there is a lack of systematic data augmentation methods: Charging stations have a short operating history (newly built stations are usually less than six months old), and the amount of training samples required by deep learning models is far from sufficient. Calculated at a 5-minute sampling interval, a charging station that has been operating for 3 months only has about 25,920 data points, while models such as LSTM and Transformer usually require more than 50,000 data points to achieve satisfactory prediction accuracy. Existing technologies either avoid this problem or only use a single enhancement strategy (such as adding Gaussian noise), which cannot cover the complex pattern changes of charging load on four time scales: impact scale, rhythm scale, periodic scale, and seasonal scale. A single enhancement strategy is prone to introducing specific biases, which in turn impairs generalization performance.

[0012] (3) Lack of adaptive coordination and fusion ratio optimization mechanism for multi-strategy data augmentation: The effectiveness of different data augmentation strategies varies significantly under different data distributions, different charging station types, and different seasons. Existing technologies lack a multi-strategy collaborative scheduling mechanism, and fixed weights or manual parameter tuning cannot achieve optimal configuration. At the same time, the optimal fusion ratio of real data and multi-source augmented data changes dynamically with data distribution, and there is a lack of adaptive optimization methods.

[0013] (4) Lack of an end-to-end closed-loop optimization chain: In existing technologies, the selection and parameter configuration of data augmentation strategies, the determination of fusion ratios, and the training and evaluation of prediction models are independent processes. The selection of augmentation strategies is usually based on experience and conventions, lacking feedback-driven adjustments based on actual prediction results. Error information from prediction models cannot be effectively transmitted to the data augmentation stage to guide strategy optimization, resulting in the system lacking the ability to continuously improve itself through running data.

[0014] In summary, existing technologies have shortcomings in areas such as unified access to multi-source heterogeneous data, solutions for scarce training data, adaptive collaboration of multi-strategy data augmentation, and automatic optimization of the real-to-expanded data ratio, and fail to provide effective technical solutions. Summary of the Invention

[0015] In view of this, the purpose of this invention is to provide a charging station load prediction data processing method and system based on dynamic fusion of multi-source heterogeneous data. In view of the scarcity of charging station data and the characteristics of multi-scale composite load, the method achieves high-precision load prediction through deep fusion of multi-source data, six-strategy collaborative sample expansion, reinforcement learning weight optimization and closed-loop feedback mechanism.

[0016] To achieve the above objectives, the present invention provides the following technical solution: On the one hand, a method for processing charging station load forecasting data based on dynamic fusion of multi-source heterogeneous data is provided, the method including: S1. Obtain historical power data, real-time meteorological data and charging pile attribute data of charging stations respectively through the pluggable data source adapter mode to form the original multi-source heterogeneous dataset; S2. Perform four cascaded processing steps on the original multi-source heterogeneous dataset: data validation, time alignment, cleaning and filtering, and normalization transformation, to output a standardized feature dataset. S3. The standardized feature dataset is augmented with multiple heterogeneous augmentation strategies in parallel. The augmented candidate samples output in parallel are then fused to obtain a unified candidate pool. In the fusion process, the strategy weights are dynamically adjusted in real time based on the prediction evaluation feedback signal to adjust the weighted contribution of each heterogeneous augmentation strategy. S4. Employ a reinforcement learning-based dynamic optimization method to fuse the obtained unified candidate pool with the standardized feature dataset in a proportional manner, and output the final fused dataset. S5. Establish a time-series forecasting model for charging station load forecasting, and use the final fused dataset to train and validate the time-series forecasting model, and output the evaluation results. S6. Feed the evaluation results back to the weighted fusion process of multiple heterogeneous expansion strategies and the dynamic optimization process of the fusion ratio based on reinforcement learning, drive the expansion strategy weights and fusion ratios to continuously iterate and optimize until the preset stopping condition is reached, and output the optimal model parameters. Use the time series prediction model under the optimal model parameters for charging station load prediction.

[0017] Furthermore, in step S1, the pluggable data source adapter pattern employs a three-tier architecture to implement data source adaptation, including an abstract base class layer, an adapter subclass layer, and a factory registry layer. The abstract base class layer defines a unified interface protocol, including connection, disconnection, data retrieval, health checks, and statistical information retrieval; The adapter subclass layer converts the original data format of the corresponding data source into a unified internal representation according to different adapter types. The adapter types include historical power data adapter, meteorological data adapter, and charging pile attribute adapter. The factory registry layer manages all adapter instances uniformly through adapter registration, providing registration and retrieval functions; The acquired historical power data includes charging power, charging capacity, number of charging vehicles, and charging time; the acquired real-time meteorological data includes temperature, humidity, wind speed, wind direction, weather type, and air pressure; the acquired charging pile attribute data is divided into static attributes and dynamic attributes. Static attributes include pile type, rated power, maximum output current, charging interface type, and geographical coordinates, while dynamic attributes include real-time operating status and current occupancy status. Each adapter performs a health check before each data acquisition, confirms the data source connection status through a heartbeat detection mechanism, and removes data source information that has timed out.

[0018] Furthermore, in step S2, the cascaded processing and standardization of the original multi-source heterogeneous dataset is as follows: First, data validation is performed, including integrity validation and compliance validation. Integrity validation verifies the identified core fields, and data missing core fields is marked as invalid and removed. Compliance validation verifies the data value range to ensure that the value range of each data item is within the preset range. Then, time alignment is performed to unify the reference time granularity and align data from different sampling frequencies to a unified timestamp: for high-frequency data, window averaging is used for downsampling; for low-frequency data, linear interpolation is used for upsampling. Next, a cleaning and filtering process is performed, using the three-standard-deviation criterion to detect outliers, and calculating the mean for each feature dimension. and standard deviation It will exceed Data points within a range are marked as anomalies, and domain rules are used to assist in anomaly detection; outliers are removed or filled in. Finally, normalization conversion is performed, using different standardization methods for different data types.

[0019] Furthermore, in step S3, the parallel heterogeneous augmentation strategies include a time window extension strategy, a multi-site generation strategy, a noise injection enhancement strategy, a model resampling strategy, a cross-seasonal migration strategy, and a combined meteorological and load augmentation strategy. For the time window expansion strategy, the sampling granularity is determined by calculating the median of the original data time interval based on the periodicity of the daily and weekly cycles of the charging load. Index by Hour and week type The mean and standard deviation of the power in the group were statistically analyzed to obtain... Distribution parameters are then used to generate augmented sample values ​​through Gaussian sampling within the target time window. ; For the multi-site generation strategy, virtual charging station data with similar electrical characteristics but different configurations are generated by adjusting parameters. The number of adoptions includes the capacity coefficient. Noise level Time offset Capacity factor Simulating noise levels at charging stations of different sizes Simulated measurement noise difference, time offset Simulate time differences across different geographical locations; For the noise injection enhancement strategy, three sub-strategies are used to enhance the local robustness of the training data, including Gaussian noise injection, temporal jitter, and power scaling. For the pattern resampling strategy, resampling is performed based on the statistical distribution pattern of the original data, extracting time features, and mapping the timestamp of each sample to hours. and week type ; Calculate the hourly power distribution for each The average power value of the combined statistics and standard deviation Construct conditional distributions that distinguish between weekdays and weekends, and model weekdays and weekends separately; For cross-seasonal migration strategies, the existing seasonal load patterns are migrated to other seasons by using seasonal power factors and peak period offsets. For the combined meteorological and load expansion strategy, a physical constraint model is introduced to generate meteorologically coordinated load data through multi-factor coupling.

[0020] Furthermore, in step S3, the weights of the augmented candidate samples output by each heterogeneous augmentation strategy in the fusion process are affected by the evaluation results output by the verification process. The weight adjustment is expressed as follows:

[0021] in, For the first In the first iteration The weights of each strategy For the first Marginal contribution score of each strategy For learning rate, Total number of strategies; The augmentation candidate samples from each heterogeneous augmentation strategy are weighted and fused according to the adjusted weights:

[0022] in, To expand the candidate set after fusion, For the first The strategy generates the first One sample, For its weight, For the first The number of samples generated by each strategy.

[0023] Furthermore, in step S4, the Deep Deterministic Policy Gradient (DDPG) algorithm is used to find the optimal fusion ratio between real samples and augmented samples, wherein: Define the state space:

[0024] in, This represents the difference in distribution between the augmented sample and the true sample, as measured by the KL divergence. , These represent the probability distributions of the augmented sample and the true sample, respectively. This indicates the expanded sample as measured by the Pearson correlation coefficient. Compared with real samples Linear correlation; Expanded sample as measured by data diversity entropy Information richness; Define the action space: the action vector is The policy weights are output after Softmax normalization:

[0025] in, The original weight vector output by the Actor network. For the first The original weight values ​​of each strategy; Define the reward function: The reward function is a weighted sum of the prediction accuracy reward and the diversity reward.

[0026] in, For the first The immediate rewards obtained from each iteration These are the weighting coefficients of the reward function. For the prediction model in the 1st The root mean square error on the validation set. The preset upper limit of error, The maximum possible entropy value, For indicator functions, when The value is 1 if the condition is met, otherwise it is 0. The error threshold for triggering diversity rewards.

[0027] Furthermore, in step S4, the process of training the defined Actor-Critic network and obtaining the final fused dataset is as follows: Actor Network Output action vectors, Critic network Evaluate the Q-value of the state-action pair; and initialize the experience replay pool capacity, random sampling batch size, reward change threshold, and total number of iterations; The Critic network updates by minimizing the TD error, and its loss function is:

[0028] in, For batch size, As a discount factor, and The target network parameters are updated using a soft update method, which consists of the target Critic network and the target Actor network. The Actor network updates via policy gradients, as shown in the formula:

[0029] The exploration strategy employs Ornstein-Uhlenbeck process noise, which decays exponentially with training. ,in The attenuation coefficient; When the reward change is less than the reward change threshold for several consecutive iterations, or when the total number of iterations is reached, training stops and the current optimal policy weight combination and fusion ratio are output. The unified candidate pool and the standardized feature dataset are fused proportionally according to the fusion ratio, and the final fused dataset is output:

[0030] in, The optimal fusion ratio is the output, representing the proportion of augmented samples in the final training set; This is a standardized feature dataset, i.e., real samples; To unify the candidate pool, that is, to expand the sample.

[0031] Furthermore, in step S5, the training process of the time-series prediction model used for charging station load prediction is as follows: The dataset is divided into training, validation, and test sets in chronological order. An LSTM neural network was used as the load forecasting model. The network parameters were initialized, including batch size, training epochs, optimizer, learning rate, and L2 regularization coefficient. The mean squared error is defined as the loss function. An early stopping strategy is adopted during training. Training is stopped when the loss on the validation set does not decrease for a preset number of consecutive rounds. The root mean square error (RMSE) and mean absolute percentage error (MAPE) are used as evaluation metrics. The decision to proceed to the next round of optimization is based on the set RMSE and MAPE thresholds. If both RMSE and MAPE are below the corresponding thresholds, the training ends; otherwise, a new round of optimization is performed, and the evaluation results are fed back to the sample fusion process and the ratio optimization process.

[0032] Furthermore, in step S6, continuous closed-loop feedback iteration is performed based on the evaluation results, including: Calculate the difference between the RMSE of the current validation set and the RMSE of the previous validation set to obtain the accuracy improvement. The formula is as follows:

[0033] in, For the first The root mean square error of the validation set of the round. This represents the root mean square error of the previous round of validation sets; positive values ​​indicate improved accuracy, while negative values ​​indicate decreased accuracy. When the improvement in accuracy is greater than a preset threshold, the confidence coefficient of the current weight combination is increased; when the improvement in accuracy is negative, an exploration mechanism is triggered to introduce random perturbations in the weight update. When the absolute value of the accuracy improvement is less than the preset convergence threshold and is continuous When the conditions are met, the closed-loop iteration stops, indicating that the system has entered steady-state operation. The system outputs the current optimal strategy weight combination and fusion ratio, as well as the optimal model parameters for the current round, and deploys the prediction model as an online prediction service.

[0034] On the other hand, a system is also provided for executing the aforementioned charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data, the system comprising: Multi-source data access module: Based on a pluggable adapter architecture, it adapts to multi-source heterogeneous data sources such as historical power, meteorological data, charging pile attributes and electricity price time period data, and supports hot-swappable registration of new data sources; Cascaded data processing module: Performs four cascaded processing steps: data validation, time alignment, cleaning and filtering, and normalization transformation, and outputs a standardized feature dataset; Multi-strategy collaborative sample augmentation module: integrates six heterogeneous augmentation strategies, including time window expansion, multi-site generation, noise injection enhancement, model resampling, cross-seasonal migration, and meteorological-load joint sample generation, and achieves strategy synergy through a strategy weight dynamic adjustment module; Reinforcement learning fusion optimization module: Based on the DDPG algorithm, it uses a three-dimensional data quality evaluation vector as the state, a combination of policy weights as the action, and a joint reward function as the optimization objective to adaptively find the optimal policy weights and fusion ratio; Load forecasting and evaluation module: Based on LSTM neural network, the load forecasting model is trained on the fused dataset and outputs RMSE and MAPE evaluation indicators; Closed-loop feedback module: Receives prediction and evaluation results and automatically feeds them back to the strategy weight adjustment module and the fusion optimization module to achieve end-to-end closed-loop continuous optimization.

[0035] The beneficial effects of this invention are as follows: First, this invention adopts a pluggable data source adapter mode, which can flexibly adapt to various heterogeneous data sources such as historical power data of charging stations, real-time meteorological data, and charging pile attribute data. This breaks through the technical barriers of traditional data processing solutions, such as single data source access, poor compatibility, and insufficient scalability. It can achieve efficient access to multi-dimensional and full-scenario charging station related data, providing a comprehensive and complete data foundation for subsequent accurate load forecasting. At the same time, the pluggable architecture facilitates the expansion of new data sources in the future, greatly improving the system's general adaptability and iterative upgrade space.

[0036] Secondly, this invention constructs a standardized cascaded data processing pipeline, which sequentially completes data verification, time alignment, impurity cleaning, and unified normalization processing. This effectively solves data quality problems such as disordered formats, misaligned time sequences, outlier interference, and inconsistent units in multi-source heterogeneous data. It completely eliminates the negative impact of invalid and erroneous data on the prediction model, significantly improves the regularity and reliability of input data, provides high-quality data support for model training, and avoids the model fitting deviation problem caused by data defects in traditional load forecasting from the source.

[0037] Most importantly, this invention establishes a multi-strategy collaborative sample expansion engine comprising six heterogeneous strategies: time window expansion, multi-site generation, noise injection enhancement, pattern resampling, cross-seasonal migration, and meteorological-load joint sample generation. This engine comprehensively expands effective training samples across time, space, scenario, and correlation dimensions, completely solving the core challenges of scarce historical data and insufficient sample coverage in scenarios such as the cold start of newly built charging stations and niche stations. It also overcomes the shortcomings of traditional single-sample expansion strategies, such as weak scenario adaptability and insufficient sample diversity. Furthermore, through a reinforcement learning-based fusion ratio optimizer, the weights of each expansion strategy and the ratio of real data can be dynamically adjusted according to the load characteristics, seasonal scenarios, and data quality of different charging stations. This eliminates the rigidity of traditional fixed-weight fusion modes, achieving adaptive dynamic optimization of the data fusion strategy and significantly improving the authenticity, effectiveness, and scenario adaptability of the expanded samples.

[0038] Finally, this invention establishes an end-to-end closed-loop feedback iteration mechanism, which can continuously optimize dataset construction, strategy weight allocation, and time-series prediction model parameters based on real-time prediction results. This enables closed-loop operation of the entire process of data processing, model training, prediction inference, and iterative optimization, allowing the system to adapt to complex dynamic conditions such as weather changes, charging traffic fluctuations, and seasonal changes, and continuously improve the stability and accuracy of model prediction.

[0039] In summary, this invention effectively solves the industry challenges of difficult multi-source data fusion for charging stations, low prediction accuracy in small sample scenarios, and difficulty in accurately fitting the multi-scale composite characteristics of loads. It greatly improves the load prediction performance of charging stations during cold starts and under all operating conditions, and can provide accurate and reliable data support for charging station operation and maintenance scheduling, grid load balancing, and energy storage configuration optimization. It has extremely high engineering promotion value and economic value.

[0040] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0041] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the overall process of the charging station load prediction data processing method based on dynamic fusion of multi-source heterogeneous data according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall architecture of charging station load prediction data processing based on dynamic fusion of multi-source heterogeneous data in an embodiment of the present invention. Detailed Implementation

[0042] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0043] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0044] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0045] Please see Figures 1-2 This is a data processing method and system for charging station load prediction based on dynamic fusion of multi-source heterogeneous data.

[0046] Example 1 This embodiment first describes the detailed implementation process of a charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data, such as... Figure 1 As shown, it includes the following steps: S1. Unified access to multi-source heterogeneous data.

[0047] This step focuses on the unified access and adaptation management of multiple heterogeneous data sources. Follow these steps: (1) Pluggable data source adapter architecture design.

[0048] A three-tier architecture is adopted to achieve data source adaptation: The abstract base class layer defines a unified interface contract, including five core methods: connect, disconnect, fetch_data, health_check, and get_statistics; The adapter subclass layer implements the historical power data adapter (PowerDataAdapter), weather data adapter (WeatherDataAdapter), and charging pile attribute adapter (PileAttributeAdapter), respectively. Each adapter subclass is responsible for converting the raw data format of a specific data source into a unified internal representation; The factory registry layer manages all adapter instances in a unified manner through the adapter registry (AdapterRegistry), providing the registration function register_adapter(source_type, adapter_class) and the retrieval function get_adapter(source_type), supporting dynamic registration and discovery at runtime.

[0049] (2) Implementation of multi-source data access.

[0050] Access the following data sources via the registered adapter: Historical power data of charging stations, supporting four time granularities: 1 min, 5 min, 15 min, and 1 h, including fields such as charging power (kW), charging capacity (kWh), number of charging vehicles, and charging duration; Real-time meteorological data, supporting multiple API providers such as the China Meteorological Administration, OpenWeatherMap, and Weather Underground, with data fields including temperature (°C), humidity (%), wind speed (m / s), wind direction, weather type (sunny / cloudy / rainy / snowy), and air pressure (hPa), with a sampling interval ≤30 min; Charging pile attribute data, including pile type (DC fast charging / AC slow charging / supercharging), rated power (kW), maximum output current (A), and charging interface type (GB / T). Static attributes such as 20234 / CCS / CHAdeMO and geographic coordinates (latitude and longitude), as well as dynamic attributes such as real-time operating status (online / offline / fault) and current occupancy status (idle / charging / reserved); electricity price time period data, collecting local time-of-use electricity price information, including the start and end times of peak hours, normal hours, and valley hours and the corresponding electricity price (yuan / kWh).

[0051] (3) Data source health monitoring and anomaly handling.

[0052] Each adapter performs a health check before each data acquisition, confirming the data source connection status through a heartbeat detection mechanism with a detection cycle of 30 seconds. If three consecutive heartbeats fail, a reconnection mechanism is automatically triggered with an incremental backoff strategy (1s, 3s, 9s, 27s). When the data source remains unavailable (for more than 5 minutes), the data source status is marked as degraded, and an alarm is issued through the log system. At the same time, the system is downgraded to using the latest cached data. When all data sources are unavailable, the system enters safe mode, suspends the data expansion process, and awaits manual intervention.

[0053] S2, Cascaded Data Processing and Standardization.

[0054] This step is executed after data access is complete. It involves performing four cascading processing steps on the multi-source raw data, as follows: (1) Data validation.

[0055] Check the completeness of required fields, including core fields such as timestamp, charging power, and ambient temperature. Records with missing required fields are marked as invalid and removed. Verify the compliance of value ranges: charging power should be non-negative and not exceed 1.2 times the rated power; SOC should be within the range of 0-100%; temperature should be within the range of -40℃-60℃; and humidity should be within the range of 0-100%. Record the validity rate of each field. Issue a data quality alarm when the validity rate of a field is below 90%.

[0056] (2) Time alignment.

[0057] A unified reference time granularity of 5 minutes is used to align data from different sampling frequencies to a unified timestamp. For high-frequency data (e.g., 1-minute sampling), window mean downsampling is employed, with the downsampling formula as follows:

[0058] in, This represents the aggregate power value over a 5-minute window. This represents the original number of sampling points within the window. For the first in the window The original sample values ​​are used; for low-frequency data (such as 15 min, 1 h), linear interpolation upsampling is used. When the interpolation ratio exceeds 30%, the time period is marked as a low confidence interval, and the weight coefficient of the data in this interval is reduced in subsequent predictions.

[0059] (3) Clean the filter.

[0060] Using the three-standard-deviation criterion ( (Criteria) Detect outliers and calculate the mean for each feature dimension. and standard deviation It will exceed Data points within a given interval are marked as outliers, using the following formula:

[0061] Anomaly detection is aided by domain-specific rules: a sudden increase in charging power exceeding 50% of the rated power for a duration ≤1 sampling period is considered a normal surge in overcharging; charging power exceeding twice the daily average power at night (0:00~6:00) is marked as a suspected anomaly requiring manual review. The missing value imputation strategy is as follows: if there are two or more valid values ​​before and after a missing value, linear interpolation of adjacent means is used; if the missing segment exceeds 5 consecutive sampling points, predicted values ​​based on a time-series prediction model (LSTM) are used for imputation, with the following formula:

[0062] in, for The fill value at time, To fill in the gaps in the pre-trained LSTM model, The length of the backtracking window (taken in the example) (corresponding to 1 hour of data).

[0063] (4) Normalization conversion.

[0064] Continuous features are scaled using Min-Max to map numerical values ​​to... The interval, the formula is:

[0065] Alternatively, Z-score standardization can be used, with the formula as follows:

[0066] in, These are the minimum and maximum values ​​of this feature in the training set, respectively. These represent the mean and standard deviation of the feature in the training set. For features with clear physical meaning, such as charging power, Min-Max scaling is used; for normally distributed features, such as temperature, Z-score standardization is used. Categorical features use one-hot encoding, and weather type (sunny / cloudy / rainy / snowy) is mapped to a 4-dimensional binary vector. Time features are decomposed into four numerical features: timestamp, hour, day of the week, whether it is a working day (0 / 1), and month (1~12). Normalization parameters (minimum, maximum, mean, and standard deviation) are persistently saved to a parameter configuration file for loading during online deployment.

[0067] S3, Multi-strategy Collaborative Sample Augmentation.

[0068] like Figure 2 As shown, this step is performed after data standardization is complete. The standardized feature dataset is input in parallel into six heterogeneous augmentation strategies to generate a multi-path augmentation candidate sample set, as follows: (1) Time window expansion strategy.

[0069] Based on the 24-hour daily cycle and 7-day weekly cycle of charging load, the median of the time interval of the raw data is calculated to determine the sampling granularity. Index by Hour and week type The mean and standard deviation of the power in the group were statistically analyzed to obtain... Distribution parameters; Gaussian sampling is performed within the target time window, using the following formula:

[0070] in, For the generated augmented sample values, For For the mean, The variance follows a Gaussian distribution. This strategy is suitable for cold start scenarios with less than 3 months of running history but a complete cycle pattern has been accumulated, and the output expanded data volume is approximately 12,500 records.

[0071] (2) Multi-site generation strategy.

[0072] Virtual charging station data with similar electrical characteristics but different configurations is generated by adjusting parameters. The core parameters include: capacity coefficient. Random sampling was performed using a uniform distribution to simulate charging stations of different sizes; noise levels Random sampling is performed according to a uniform distribution to simulate and measure noise differences; time offset Random sampling is performed using a uniform distribution to simulate time differences across different geographical locations. The formula for generating virtual station data is:

[0073] in, For virtual site One sample, This is the sample corresponding to the original station. To amplify noise, this strategy outputs approximately 10,000 additional data entries.

[0074] (3) Noise injection enhancement strategy.

[0075] To enhance the local robustness of training data, three sub-strategies are employed, allocated in a 1:1:1 ratio: Gaussian noise injection, adding data with a mean of 0 and a standard deviation of... (from Gaussian noise (uniformly sampled from intervals), the formula is: Time jitter, applied to each sample on the timeline. A random offset of min is applied, keeping the charging power value constant while changing the time label of the sample; power scaling is applied to the charging power value. The random scaling factor for the interval simulates the random error in power measurement. This strategy outputs approximately 7500 augmented data entries.

[0076] (4) Pattern resampling strategy.

[0077] Resampling based on the statistical distribution pattern of the original data involves the following steps: extracting time features and mapping the timestamp of each sample to hours. and week type ; Calculate the hourly power distribution for each The average power value of the combined statistics and standard deviation ; Construct conditional distributions that distinguish between weekdays and weekends, modeling weekdays (Monday to Friday) and weekends (Saturday and Sunday) separately; Prioritize sampling based on refined patterns, using the following formula:

[0078] and through This strategy ensures the non-negativity of the generated samples. The output augmented data is approximately 9000 records.

[0079] (5) Cross-seasonal migration strategy.

[0080] The existing seasonal load pattern is migrated to other seasons by using the seasonal power factor and peak period offset. The seasonal power factor value is: Spring (March-May). Summer (June-August) Autumn (September-November) Winter (December-February) The peak period shift is as follows: the summer peak period shifts forward by 1 hour (due to higher temperatures causing earlier charging peaks), and the winter peak period shifts backward by 1 hour (due to lower temperatures causing later charging peaks). The migration formula is:

[0081] in, For the target season Migration power at time, For seasonal power factor, This represents the actual power of the source season after time offset. This strategy outputs approximately 6000 augmented data entries.

[0082] (6) Meteorological-load joint expansion strategy.

[0083] A physical constraint model is introduced, and meteorologically coordinated load data are generated through multi-factor coupling. The formula for the annual temperature cycle model is:

[0084] in, The first of the year The base temperature of the day The average annual temperature (taken in the example) ), Annual temperature amplitude (taken in the example) ), The daily shift of the lowest temperature (take) (corresponding to around January 20th).

[0085] The formula for the daily temperature cycle model is:

[0086] in, The first of the day Hourly temperature fluctuations, The daily temperature amplitude (taken in the example) ), The hour in which the highest temperature occurs (14 in this example, corresponding to 2 PM).

[0087] The humidity-temperature coupling relationship is described by a linear regression model: ,in Relative humidity (%) Temperature (°C) These are regression coefficients. Weather condition influence coefficients are: sunny 1.05, cloudy 0.95, rain 0.85, snow 0.70. Temperature influence factor values ​​are: high temperature (…). Take 1.25, low temperature ( We take 1.20 for the normal temperature range and 1.00 for the normal temperature range. The final formula for the meteorologically coordinated load data is:

[0088] in, For the generated load data, Based on the load, The influence coefficient of weather conditions. Temperature is a factor that affects the environment. This is for coupled noise. This strategy outputs approximately 5000 augmented data entries.

[0089] (7) Dynamic adjustment and integration of strategy weights.

[0090] The strategy weight dynamic adjustment module adjusts the weighted contribution of each strategy in real time based on the prediction and evaluation feedback signals. The weight adjustment formula is as follows:

[0091] in, For the first In the first iteration The weights of each strategy For the first The marginal contribution score of each strategy (based on the improvement in RMSE on the validation set using augmented data of that strategy). The learning rate (taken in the example) ), The total number of strategies (in this embodiment, we take...) After the weights are normalized using Softmax, the augmented candidate samples from each path are weighted and fused according to their respective weight ratios. The formula is as follows:

[0092] in, To expand the candidate set after fusion, For the first The strategy generates the first One sample, For its weight, For the first The number of samples generated by each strategy. The lower limit of the weight is set to 0.05 to avoid completely excluding a certain strategy.

[0093] S4. Optimize the integration ratio of reinforcement learning.

[0094] This step, based on the dynamic adjustment of policy weights, uses the Deep Deterministic Policy Gradient (DDPG) algorithm to automatically find the optimal fusion ratio between real and augmented samples. The operation is as follows: (1) Definition of state space.

[0095] The state space is a three-dimensional data quality assessment vector, and the formula is:

[0096] The KL divergence measures the difference in distribution between the augmented sample and the true sample, and its formula is:

[0097] , They represent the probability distributions of the augmented sample and the true sample, respectively. The augmented sample and the real sample respectively fall into the first The probability of each power bin. The total number of boxes (in the example, it is taken as) The Pearson correlation coefficient measures the linear correlation between the augmented sample and the original sample, and its formula is:

[0098] Data diversity entropy measures the information richness of augmented samples, and the formula is:

[0099] (2) Definition of action space.

[0100] Action vector is The policy weights, after being normalized by Softmax, are output using the following formula:

[0101] in, The original weight vector output by the Actor network. For the first The original weight values ​​of each strategy.

[0102] (3) Reward function design.

[0103] The joint reward function is a weighted sum of the prediction accuracy reward and the diversity reward, and the formula is:

[0104] in, For the first The immediate rewards obtained from each iteration The weight coefficients of the reward function (satisfying) In the example, take ), For the prediction model in the 1st The root mean square error on the validation set. The preset upper limit of error (taken in the example) ), To expand the data diversity entropy of the candidate set, The maximum possible entropy value, For indicator functions (when) (Take 1 if the condition is met, otherwise take 0). To trigger the error threshold for diversity rewards (in the example, it is taken as...) ).

[0105] (4) DDPG algorithm training process.

[0106] Employing a dual-network architecture of Actor-Critic, the Actor network... Output action vectors, Critic network The Q-value of the state-action pair is evaluated; the empirical replay pool capacity is set to 10000, and each training iteration randomly samples transitions with a batch size of 64; the Critic network updates by minimizing the TD error, with the loss function being:

[0107] in, For batch size, This is the discount factor (0.95 in this example). and The target network and the target actor network are respectively defined. The parameters of the target network are updated using a soft update method, and the soft update coefficients are... .

[0108] The Actor network updates via policy gradients, as shown in the formula:

[0109] The exploration strategy employs Ornstein-Uhlenbeck process noise, with an initial exploration rate of 0.3 that decays exponentially with training. ,in This is the attenuation coefficient (0.01 in this example).

[0110] (5) Conditions for termination of training.

[0111] When the reward change is less than 0.01 for 10 consecutive iterations, or when the total number of iterations reaches 200, RL training is stopped, the current optimal policy weight combination and fusion ratio are output, and the optimal fusion ratio is fixed and used for subsequent prediction model training.

[0112] (6) Construction of fusion dataset.

[0113] Expanded candidate set after adjusting strategy weights Compared with the real sample set According to the optimal fusion ratio output by the RL optimizer Merge and construct the final training dataset, using the following formula:

[0114] in, The optimal fusion ratio (typical values ​​in the example) (between), indicating the proportion of the expanded samples in the final training set.

[0115] S5. Load forecasting model training and evaluation.

[0116] This step is performed after the optimal weights are output through RL optimization. The load prediction model is trained using the fused dataset, and the operation is as follows: (1) Data set partitioning.

[0117] The dataset is divided into a training set (first 70%), a validation set (middle 15%), and a test set (last 15%) in chronological order to ensure that the data is not leaked.

[0118] (2) Prediction model training.

[0119] An LSTM neural network was used as the load prediction model. The network structure consisted of two LSTM layers with a hidden layer dimension of 128, followed by a fully connected layer to output the predicted value. Training parameters included a batch size of 64, 200 training epochs, an Adam optimizer, a learning rate of 0.001, and an L2 regularization coefficient. The loss function uses mean squared error (MSE), and the formula is as follows:

[0120] in, For the first The measured power value of each sample. To predict power values ​​for the model, For the set of model parameters, is the regularization coefficient. An early stopping strategy is used during training; training stops when the validation set loss does not decrease for 10 consecutive rounds.

[0121] (3) Calculation of evaluation indicators.

[0122] The root mean square error (RMSE) and mean absolute percentage error (MAPE) are used as the main evaluation indicators, and the formulas are as follows:

[0123]

[0124] in, To assess sample size. When MAPE is below 10% and RMSE is below If the model meets the deployment requirements, it is determined that the model meets the requirements; otherwise, a new round of RL fusion optimization is triggered to adjust the strategy weights and fusion ratios.

[0125] S6, End-to-end closed-loop feedback iteration.

[0126] This step covers the entire process of data augmentation, fusion, training, and evaluation, and is performed as follows: (1) Calculation of accuracy improvement.

[0127] Calculate the difference between the RMSE of the current validation set and the RMSE of the previous validation set to obtain the accuracy improvement. The formula is as follows:

[0128] in, For the first The root mean square error of the validation set of the round. This represents the root mean square error of the previous validation set; a positive value indicates improved accuracy, while a negative value indicates decreased accuracy.

[0129] (2) Weight reset and reliability update.

[0130] When the accuracy improvement is greater than a preset threshold (taken in the example), ,Right now When this happens, the confidence coefficient of the current weight combination is increased. The confidence update formula is:

[0131] in, The confidence increment (taken in the example) When the accuracy improvement is negative, an exploration mechanism is triggered, introducing a random perturbation into the weight update. The perturbation magnitude is proportional to the accuracy decrease, as shown in the formula:

[0132] in, The mean is Standard deviation is Gaussian noise, Based on the amplitude of the disturbance (taken in the example) ), For the first The strategy weight vector of the round, To explore candidate weight vectors after perturbation, they are then normalized using Softmax and combinations that improve accuracy are retained.

[0133] (3) Iteration termination condition.

[0134] When the absolute value of the accuracy improvement is less than the preset convergence threshold (taken in the example), And continuous Wheel (taken in the example) When the conditions are met, stop the closed-loop iteration, determine that the system has entered steady-state operation, output the current optimal strategy weight combination and fusion ratio, and deploy the prediction model as an online prediction service. The convergence determination formula is:

[0135] (4) Steady-state monitoring and re-optimization.

[0136] After the system reaches a steady state, it enters a low-frequency monitoring mode, performing a lightweight re-evaluation every 7 days: calculating the MAPE predicted online for the most recent 7 days and comparing it with the baseline MAPE at the time of deployment; when the relative increase in MAPE exceeds 10%, a complete re-optimization process is triggered, including data re-collection, policy weight readjustment, and RL fusion optimization retraining; when there are significant changes in the data source or charging station configuration (such as adding a new type of charging pile or replacing the metering equipment), a full re-optimization is triggered immediately.

[0137] Example 2 This embodiment provides a system for executing the charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data in Embodiment 1, such as... Figure 2 As shown, the system is structured with a data source layer, adapter layer, processing layer, extended optimization layer, and prediction model layer. Specifically, the system comprises six functional modules that work collaboratively: Multi-source data access module: Based on a pluggable adapter architecture, it adapts to multi-source heterogeneous data sources such as historical power, meteorological data, charging pile attributes and electricity price time period data, and supports hot-swappable registration of new data sources.

[0138] The cascaded data processing module performs four cascaded processing steps: data validation, time alignment, cleaning and filtering, and normalization transformation, and outputs a standardized feature dataset.

[0139] Multi-strategy collaborative sample augmentation module: integrates six heterogeneous augmentation strategies, including time window extension, multi-site generation, noise injection enhancement, model resampling, cross-seasonal migration, and meteorological-load joint sample generation, and achieves strategy synergy through a strategy weight dynamic adjustment module.

[0140] The reinforcement learning fusion optimization module is based on the DDPG algorithm. It uses a three-dimensional data quality evaluation vector as the state, a combination of policy weights as the action, and a joint reward function as the optimization objective to adaptively find the optimal policy weights and fusion ratio.

[0141] Load forecasting and evaluation module: Based on LSTM neural network, the load forecasting model is trained on the fused dataset and outputs RMSE and MAPE evaluation indicators.

[0142] Closed-loop feedback module: Receives prediction and evaluation results and automatically feeds them back to the strategy weight adjustment module and the fusion optimization module to achieve end-to-end closed-loop continuous optimization.

[0143] Example 3 This embodiment, based on the charging station load prediction data processing method and system provided in Embodiments 1 and 2, conducts tests and verifications in three typical scenarios to illustrate the practical effects of the invention; the specific implementation scenarios and experimental results are as follows: (1) Cold start scenario test of a single charging station.

[0144] The test subject was a charging station in a city's commercial center, located in the underground parking lot. It was equipped with 12 120kW DC fast charging piles and 4 7kW AC slow charging piles, serving approximately 200 vehicles per day. The charging station began operation in January 2025, initially accumulating only two months' worth of historical load data (February to March 2025), with a raw data volume of 576 records (sampling interval of 5 minutes). A six-strategy collaborative expansion engine expanded the data to approximately 50,000 records, an expansion factor of approximately 87 times. The output data volumes for each strategy were as follows: time window expansion 12,500 records, multi-site generation 10,000 records, noise injection enhancement 7,500 records, model resampling 9,000 records, cross-seasonal migration 6,000 records, and meteorological-load joint expansion 5,000 records. The reinforcement learning fusion optimizer uses the DDPG algorithm and converges after approximately 200 iterations. The optimal weight distribution for the six strategies is as follows: time window expansion 0.25, multi-site generation 0.20, noise injection enhancement 0.10, pattern resampling 0.18, cross-seasonal transfer 0.15, and weather-load joint augmentation 0.12. An LSTM prediction model (2 layers of LSTM plus 1 fully connected layer, hidden layer dimension 128, 200 training epochs, batch size 64, Adam optimizer learning rate 0.001) is trained on the augmented dataset. The RMSE on the validation set is reduced by approximately 35% compared to training with only the original data (from 32.5kW to 21.1kW), and by approximately 18% compared to using a single strategy augmentation (time window expansion only) (from 25.7kW to 21.1kW). In a 14-day online test, the daily load forecast MAPE was kept below 8% (7.2% on weekdays and 8.5% on weekends).

[0145] (2) Multi-site generalization scenario test.

[0146] The experiment was conducted at three charging stations in different geographical locations: Station A is located in the city's central business district, equipped with 24 60kW fast charging piles, serving an average of 350 vehicles per day, with a peak-to-valley load ratio of 4.5:1 and 15,000 historical data entries; Station B is located in a suburban residential area, equipped with 8 120kW fast charging piles and 4 7kW slow charging piles, serving an average of 80 vehicles per day, with a peak-to-valley load ratio of 2.5:1 and 5,000 historical data entries; Station C is located in a highway service area, equipped with 16 180kW supercharging piles and 2 60kW fast charging piles, serving an average of 500 vehicles per day on holidays, with a peak-to-valley load ratio of up to 6:1 and 3,000 historical data entries. The RL fusion optimizer adaptively adjusts the weight allocation for each station: Station A (with sufficient data) has the highest weight for the multi-site generation strategy (0.28) and the lowest weight for noise injection (0.08); Station B (with moderate data) has a combined weight of 0.35 for cross-seasonal migration and meteorological joint augmentation; Station C (with scarce data and significant impact from holidays) has the highest weight for time window expansion and model resampling (total weight of 0.52). After applying the six-strategy augmentation, the MAPE of the three stations decreased from 12.3%, 18.7%, and 22.5% to 8.2%, 9.5%, and 10.8%, respectively.

[0147] (3) Extreme weather conditions test.

[0148] During heat waves (five consecutive days with temperatures exceeding 38°C), charging load increased by 22.3%, with the MAPE of the model in this invention being 9.2%, compared to 15.7% without meteorological augmentation. During cold waves (three consecutive days with temperatures below 0°C), charging load increased by 18.5%, with the MAPE of the model in this invention being 10.1%, compared to 17.3% without meteorological augmentation. The KL divergence of the data generated by the physical constraint method is 0.12, significantly better than the 0.37 of the purely statistical method.

[0149] Test results show that the method of the present invention achieves good prediction accuracy under charging station cold start, multi-site generalization and extreme weather conditions, and can effectively cope with the challenges of scarce training samples and heterogeneous multi-source data.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for processing charging station load forecasting data based on dynamic fusion of multi-source heterogeneous data, characterized in that: The method includes: S1. Obtain historical power data, real-time meteorological data and charging pile attribute data of charging stations respectively through the pluggable data source adapter mode to form the original multi-source heterogeneous dataset; S2. Perform four cascaded processing steps on the original multi-source heterogeneous dataset: data validation, time alignment, cleaning and filtering, and normalization transformation, to output a standardized feature dataset. S3. The standardized feature dataset is augmented with multiple heterogeneous augmentation strategies in parallel. The augmented candidate samples output in parallel are then fused to obtain a unified candidate pool. In the fusion process, the strategy weights are dynamically adjusted in real time based on the prediction evaluation feedback signal to adjust the weighted contribution of each heterogeneous augmentation strategy. S4. Employ a reinforcement learning-based dynamic optimization method to fuse the obtained unified candidate pool with the standardized feature dataset in a proportional manner, and output the final fused dataset. S5. Establish a time-series forecasting model for charging station load forecasting, and use the final fused dataset to train and validate the time-series forecasting model, and output the evaluation results. S6. Feed the evaluation results back to the weighted fusion process of multiple heterogeneous expansion strategies and the dynamic optimization process of the fusion ratio based on reinforcement learning, drive the expansion strategy weights and fusion ratios to continuously iterate and optimize until the preset stopping condition is reached, and output the optimal model parameters. Use the time series prediction model under the optimal model parameters for charging station load prediction.

2. The charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data according to claim 1, characterized in that: In step S1, the pluggable data source adapter pattern uses a three-tier architecture to implement data source adaptation, including an abstract base class layer, an adapter subclass layer, and a factory registry layer. The abstract base class layer defines a unified interface protocol, including connection, disconnection, data retrieval, health checks, and statistical information retrieval; The adapter subclass layer converts the original data format of the corresponding data source into a unified internal representation according to different adapter types. The adapter types include historical power data adapter, meteorological data adapter, and charging pile attribute adapter. The factory registry layer manages all adapter instances uniformly through adapter registration, providing registration and retrieval functions; The acquired historical power data includes charging power, charging capacity, number of charging vehicles, and charging time; the acquired real-time meteorological data includes temperature, humidity, wind speed, wind direction, weather type, and air pressure; the acquired charging pile attribute data is divided into static attributes and dynamic attributes. Static attributes include pile type, rated power, maximum output current, charging interface type, and geographical coordinates, while dynamic attributes include real-time operating status and current occupancy status. Each adapter performs a health check before each data acquisition, confirms the data source connection status through a heartbeat detection mechanism, and removes data source information that has timed out.

3. The charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data according to claim 1, characterized in that: In step S2, the cascaded processing and standardization of the original multi-source heterogeneous datasets are as follows: First, data validation is performed, including integrity validation and compliance validation. Integrity validation verifies the identified core fields, and data missing core fields is marked as invalid and removed. Compliance validation verifies the data value range to ensure that the value range of each data item is within the preset range. Then, time alignment is performed to unify the reference time granularity and align data with different sampling frequencies to a unified timestamp: for high-frequency data, window averaging is used for downsampling. For low-frequency data, linear interpolation is used for upsampling; Next, a cleaning and filtering process is performed, using the three-standard-deviation criterion to detect outliers, and calculating the mean for each feature dimension. and standard deviation It will exceed Data points within a range are marked as anomalies, and domain rules are used to assist in anomaly detection; outliers are removed or filled in. Finally, normalization conversion is performed, using different standardization methods for different data types.

4. The charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data according to claim 1, characterized in that: In step S3, the parallel heterogeneous augmentation strategies include a time window extension strategy, a multi-site generation strategy, a noise injection enhancement strategy, a model resampling strategy, a cross-seasonal migration strategy, and a combined meteorological and load augmentation strategy. For the time window expansion strategy, the sampling granularity is determined by calculating the median of the original data time interval based on the periodicity of the daily and weekly cycles of the charging load. Index by Hour and week type The mean and standard deviation of the power in the group were statistically analyzed to obtain... Distribution parameters are then used to generate augmented sample values ​​through Gaussian sampling within the target time window. ; For the multi-site generation strategy, virtual charging station data with similar electrical characteristics but different configurations are generated by adjusting parameters. The number of adoptions includes the capacity coefficient. Noise level Time offset capacity factor Simulating noise levels at charging stations of different sizes Simulated measurement noise difference, time offset Simulate time differences across different geographical locations; For the noise injection enhancement strategy, three sub-strategies are used to enhance the local robustness of the training data, including Gaussian noise injection, temporal jitter, and power scaling. For the pattern resampling strategy, resampling is performed based on the statistical distribution pattern of the original data, extracting time features, and mapping the timestamp of each sample to hours. and week type ; Calculate the hourly power distribution for each The average power value of the combined statistics and standard deviation Construct conditional distributions that distinguish between weekdays and weekends, and model weekdays and weekends separately; For cross-seasonal migration strategies, the existing seasonal load patterns are migrated to other seasons by using seasonal power factors and peak period offsets. For the combined meteorological and load expansion strategy, a physical constraint model is introduced to generate meteorologically coordinated load data through multi-factor coupling.

5. The charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data according to claim 4, characterized in that: In step S3, the weights of the augmented candidate samples output by each heterogeneous augmentation strategy in the fusion process are affected by the evaluation results output by the verification process. The weight adjustment is expressed as follows: in, For the first In the first iteration The weights of each strategy For the first Marginal contribution score of each strategy For learning rate, Total number of strategies; The augmentation candidate samples from each heterogeneous augmentation strategy are weighted and fused according to the adjusted weights: in, To expand the candidate set after fusion, For the first The strategy generates the first One sample, For its weight, For the first The number of samples generated by each strategy.

6. The charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data according to claim 1, characterized in that: In step S4, the Deep Deterministic Policy Gradient (DDPG) algorithm is used to find the optimal fusion ratio between real samples and augmented samples, where: Define the state space: in, This represents the difference in distribution between the augmented sample and the true sample, as measured by the KL divergence. , These represent the probability distributions of the augmented sample and the true sample, respectively. This indicates the expanded sample as measured by the Pearson correlation coefficient. Compared with real samples Linear correlation; Expanded sample as measured by data diversity entropy Information richness; Define the action space: the action vector is The policy weights are output after Softmax normalization: in, The original weight vector output by the Actor network. For the first The original weight values ​​of each strategy; Define the reward function: The reward function is a weighted sum of the prediction accuracy reward and the diversity reward. in, For the first The immediate rewards obtained from each iteration These are the weighting coefficients of the reward function. For the prediction model in the 1st The root mean square error on the validation set. The preset upper limit of error, The maximum possible entropy value, For indicator functions, when The value is 1 if the condition is met, otherwise it is 0. An error threshold for triggering diversity rewards.

7. The charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data according to claim 6, characterized in that: In step S4, the process of training the defined Actor-Critic network and obtaining the final fused dataset is as follows: Actor Network Output action vectors, Critic network Evaluate the Q-value of the state-action pair; It also initializes the experience replay pool capacity, random sampling batch size, reward change threshold, and total number of iterations; The Critic network updates by minimizing the TD error, and its loss function is: in, For batch size, As a discount factor, and The target network parameters are updated using a soft update method, which consists of the target Critic network and the target Actor network. The Actor network updates via policy gradients, as shown in the formula: The exploration strategy employs Ornstein-Uhlenbeck process noise, which decays exponentially with training. ,in The attenuation coefficient; When the reward change is less than the reward change threshold for several consecutive iterations, or when the total number of iterations is reached, training stops and the current optimal policy weight combination and fusion ratio are output. The unified candidate pool and the standardized feature dataset are fused proportionally according to the fusion ratio, and the final fused dataset is output: in, The optimal fusion ratio is the output, representing the proportion of augmented samples in the final training set; This is a standardized feature dataset, i.e., real samples; To unify the candidate pool, that is, to expand the sample.

8. The charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data according to claim 1, characterized in that: In step S5, the training process of the time-series prediction model used for charging station load prediction is as follows: The dataset is divided into training, validation, and test sets in chronological order. An LSTM neural network was used as the load forecasting model. The network parameters were initialized, including batch size, training epochs, optimizer, learning rate, and L2 regularization coefficient. The mean squared error is defined as the loss function. An early stopping strategy is adopted during training. Training is stopped when the loss on the validation set does not decrease for a preset number of consecutive rounds. The root mean square error (RMSE) and mean absolute percentage error (MAPE) are used as evaluation metrics. The decision to proceed to the next round of optimization is based on the set RMSE and MAPE thresholds. If both RMSE and MAPE are lower than the corresponding thresholds, the training ends. Otherwise, a new round of optimization will be conducted, and the evaluation results will be fed back to the sample fusion process and the proportion optimization process.

9. The charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data according to claim 1, characterized in that: In step S6, continuous closed-loop feedback iteration is performed based on the evaluation results, including: Calculate the difference between the RMSE of the current validation set and the RMSE of the previous validation set to obtain the accuracy improvement. The formula is as follows: in, For the first The root mean square error of the validation set of the round. This represents the root mean square error of the validation set from the previous round; positive values ​​indicate improved accuracy, while negative values ​​indicate decreased accuracy. When the improvement in accuracy is greater than a preset threshold, the confidence coefficient of the current weight combination is increased; when the improvement in accuracy is negative, an exploration mechanism is triggered to introduce random perturbations in the weight update. When the absolute value of the accuracy improvement is less than the preset convergence threshold and is continuous When the conditions are met, the closed-loop iteration stops, indicating that the system has entered steady-state operation. The system outputs the current optimal strategy weight combination and fusion ratio, as well as the optimal model parameters for the current round, and deploys the prediction model as an online prediction service.

10. A system for executing the charging station load forecasting data processing method based on dynamic fusion of multi-source heterogeneous data as described in any one of claims 1-9, characterized in that: The system includes: Multi-source data access module: Based on a pluggable adapter architecture, it adapts to multi-source heterogeneous data sources such as historical power, meteorological data, charging pile attributes and electricity price time period data, and supports hot-swappable registration of new data sources; Cascaded data processing module: Performs four cascaded processing steps: data validation, time alignment, cleaning and filtering, and normalization transformation, and outputs a standardized feature dataset; Multi-strategy collaborative sample augmentation module: integrates six heterogeneous augmentation strategies, including time window expansion, multi-site generation, noise injection enhancement, model resampling, cross-seasonal migration, and meteorological-load joint sample generation, and achieves strategy synergy through a strategy weight dynamic adjustment module; Reinforcement learning fusion optimization module: Based on the DDPG algorithm, it uses a three-dimensional data quality evaluation vector as the state, a combination of policy weights as the action, and a joint reward function as the optimization objective to adaptively find the optimal policy weights and fusion ratio; Load forecasting and evaluation module: Based on LSTM neural network, the load forecasting model is trained on the fused dataset and outputs RMSE and MAPE evaluation indicators; Closed-loop feedback module: Receives prediction and evaluation results and automatically feeds them back to the strategy weight adjustment module and the fusion optimization module to achieve end-to-end closed-loop continuous optimization.

Images