Charging load dynamic balancing method based on time series prediction algorithm

Abstract

This invention discloses a dynamic charging load balancing method based on a time-series prediction algorithm, belonging to the field of smart grid technology. The method addresses the prediction and control challenges caused by the inability of fixed time windows to adapt to multi-time-scale load fluctuations. It obtains load fluctuation patterns through adaptive clustering and analyzes short-term sensitivity indicators using sliding window technology, then dynamically adjusts the window length by integrating user behavior data. A stability correction coefficient is determined based on long-term trend data to optimize multi-time-scale prediction inputs. The intelligent dispatch system iteratively updates load balancing parameters based on seasonal deviation analysis and real-time grid feedback, forming a basis for resource allocation, and ultimately generates an energy-saving-oriented demand control sequence by classifying off-peak waste risks. This invention achieves adaptive prediction and closed-loop balancing of multi-scale charging load fluctuations, significantly improving the dispatch accuracy, resource efficiency, and energy-saving effect of the smart grid.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and more specifically to a dynamic balancing method for charging load based on time-series prediction algorithms. Background Technology

[0002] With the widespread adoption of electric vehicles, the concentrated, random, and intermittent nature of charging behavior in time and space poses significant challenges to the power balance, power quality, and operational safety of the power distribution network. Achieving accurate prediction and dynamic balancing of charging load is a key technological step in improving the grid's absorption capacity, optimizing resource allocation, and promoting the green energy transition.

[0003] Currently, existing technical solutions for charging load forecasting mainly rely on traditional time series analysis models (such as ARIMA and exponential smoothing) or machine learning methods (such as support vector machines and neural networks). These methods are typically based on fixed historical time windows and suffer from the following progressive technical bottlenecks:

[0004] At the feature extraction level, most methods rely solely on historical load data itself or simply introduce limited external factors (such as temperature and date type), failing to systematically integrate multi-source heterogeneous data such as user behavior patterns and real-time grid status. This results in insufficient model characterization of the intrinsic driving mechanism of load fluctuations and limited completeness of feature representation.

[0005] At the time-scale processing level, fixed-length analysis windows cannot adaptively capture the complex dynamic characteristics of multi-time-scale coupling in load sequences. Short-term (e.g., minute-level, hour-level) dramatic fluctuations are intertwined with long-term (e.g., daily-level, seasonal) regular trends. Existing methods, if the window is too long, smooth out crucial short-term abrupt changes, failing to respond promptly to sudden charging demands; if the window is too short, it is difficult to effectively learn and utilize long-term periodic and trend patterns, resulting in unstable prediction results.

[0006] At the level of model adaptation and closed-loop control, existing solutions mostly adopt a static mode of "offline training and online prediction," lacking a real-time feedback mechanism based on prediction deviations and grid operation status. When encountering unforeseen scenarios such as special weather, holidays, or sudden changes in user behavior patterns, the model cannot dynamically adjust its core parameters (such as window length and feature weights), and prediction deviations will continue to accumulate. Furthermore, it is difficult to effectively transform the prediction results into executable and evaluable grid dispatch instructions, thus failing to form a decision-making closed loop from "precise perception" to "dynamic equilibrium."

[0007] Therefore, how to design a charging load prediction and balancing method that can adapt to multi-timescale characteristics, integrate multi-source information for dynamic risk assessment, and optimize based on real-time feedback closed loop, fundamentally solving the rigidity of fixed window models, the limitations of single data sources, and the disconnect between open-loop prediction and closed-loop control, has become a technical challenge that urgently needs to be overcome in this field. Summary of the Invention

[0008] This invention discloses a dynamic balancing method for charging load based on a time-series prediction algorithm, which solves the technical problem that a fixed time window cannot adapt to the load fluctuation characteristics of multiple time scales, resulting in insufficient prediction accuracy.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] This invention provides a dynamic charging load balancing method based on a time-series prediction algorithm, which mainly includes:

[0011] Real-time load sequences and external environmental factors are obtained from historical charging data. Similar time period characteristics are grouped using an adaptive clustering method to obtain a preliminary classification of fluctuation patterns.

[0012] Based on the preliminary fluctuation pattern classification, the sliding window technique is used to scan continuous data segments in the sequence to determine the short-term sensitivity index in the current period.

[0013] If the short-term sensitivity index exceeds the preset threshold, user behavior data will be integrated through the information processing stage to determine whether there is a risk of sudden change and to obtain the adjusted window length range.

[0014] Long-term trend data is obtained from the adjusted window length range, and historical patterns are regrouped using an adaptive clustering method to determine the long-term stability correction coefficient.

[0015] The real-time load sequence is processed by the long-term stability correction coefficient to obtain the optimized multi-timescale prediction model input;

[0016] If the input deviation of the optimized multi-timescale prediction model increases, the information processing stage is used to analyze seasonal variation data, determine the priority of regulation needs, and obtain the final window dynamic configuration scheme.

[0017] Based on the final window dynamic configuration scheme, obtain power grid operation status feedback, and iteratively update load balance parameters through sliding window technology to determine the basis for resource efficiency allocation;

[0018] Peak load data is obtained from resource efficiency allocation criteria, and an adaptive clustering method is used to classify the risk of waste during off-peak periods, resulting in a demand regulation sequence under green transformation.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] This invention extracts real-time load sequences and external environmental factors from historical charging data, and uses an adaptive clustering method to group similar time-period features to obtain a preliminary fluctuation pattern classification. Based on this, a sliding window is used to scan continuous data segments to calculate a short-term sensitivity index. When the index exceeds a threshold, user behavior data is integrated to assess the risk of sudden changes and dynamically adjust the window length. Subsequently, long-term trend data is extracted from the adjusted window, and another adaptive clustering method is used to determine the long-term stability correction coefficient, correcting the real-time load sequence to generate an optimized multi-time-scale prediction model input. If the deviation increases, seasonal variation data is further analyzed to determine the priority of regulation demand, forming a final dynamic window configuration scheme. Finally, load balance parameters are iteratively updated based on grid operation status feedback, and peak and valley characteristics are clustered and classified to output a demand regulation sequence under green transformation. This invention effectively solves the problems of drastic fluctuations in electric vehicle charging load, large prediction deviations, and peak-valley resource mismatch leading to grid pressure and energy waste through dual adaptive adjustment of short-term sensitivity and long-term stability, and dynamic closed-loop optimization of window length. It significantly improves load prediction accuracy, resource utilization efficiency, and green and low-carbon regulation capabilities. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is an overall architecture diagram of the system described in this invention.

[0023] Figure 2 This is an overall flowchart of the method described in this invention.

[0024] Figure 3 This is a flowchart of the closed-loop optimization feedback process of the present invention. Detailed Implementation

[0025] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0026] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] Example 1: See Figures 1-3 This embodiment discloses a charging load dynamic balancing system based on a time-series prediction algorithm, the system comprising:

[0028] The data acquisition and pattern classification module is used to obtain real-time load sequences and external environmental factors from historical charging data, and to obtain preliminary fluctuation pattern classification by grouping similar time period features through an adaptive clustering method.

[0029] The short-term sensitivity analysis module, connected to the data acquisition and pattern classification module, is used to determine the short-term sensitivity index in the current period by scanning continuous data segments in the sequence using sliding window technology based on the preliminary fluctuation pattern classification.

[0030] The dynamic window adjustment module, connected to the short-term sensitivity analysis module, is used to integrate user behavior data to determine whether there is a risk of sudden change when the short-term sensitivity index exceeds a preset threshold, and generate an adjusted window length range accordingly.

[0031] The long-term stability correction module, connected to the dynamic window adjustment module, is used to obtain long-term trend data from the adjusted window length range, regroup historical patterns using an adaptive clustering method, and determine the long-term stability correction coefficient.

[0032] A multi-timescale prediction input optimization module is connected to the long-term stability correction module and is used to process the real-time load sequence through the long-term stability correction coefficient to obtain the optimized multi-timescale prediction model input.

[0033] The regulation strategy generation module is connected to the multi-timescale prediction input optimization module. It is used to analyze seasonal change data to determine the priority of regulation needs and generate the final window dynamic configuration scheme when the input display deviation of the optimized multi-timescale prediction model increases.

[0034] The load parameter iteration and resource allocation module is connected to the control strategy generation module. It is used to obtain power grid operation status feedback according to the final window dynamic configuration scheme, and to iteratively update the load balance parameters through sliding window technology to determine the basis for resource efficiency allocation.

[0035] The green regulation sequence generation module is connected to the load parameter iteration and resource allocation module. It is used to obtain peak load data from the resource efficiency allocation basis and use an adaptive clustering method to classify the waste risk during off-peak periods to obtain the demand regulation sequence under green transformation.

[0036] The dynamic balancing method for charging load based on time-series prediction algorithms includes the following:

[0037] S101: Obtain real-time load sequences and external environmental factors from historical charging data, and group similar time period features using an adaptive clustering method to obtain a preliminary fluctuation pattern classification.

[0038] Real-time load sequences and external environmental factors are obtained from historical charging data. The real-time load sequences are divided into multiple continuous time-segment sequences using a time-series segmentation method. For each continuous time-segment sequence, the average load, peak load, and load change rate are extracted as load feature vectors. External environmental factor values ​​corresponding to each continuous time-segment sequence are collected to form environmental feature vectors. The load feature vectors and environmental feature vectors are concatenated to obtain the comprehensive feature vector for each time-segment. An adaptive clustering method is used to group all comprehensive feature vectors, obtaining cluster centers and cluster labels for each time-segment. All time-segment sequences are classified according to the cluster labels to determine the preliminary fluctuation pattern classification.

[0039] In practical implementation, when processing historical charging data to obtain real-time load sequences and external environmental factors, the charging records of the past year are first extracted from the smart charging piles through the data acquisition system. For example, load data is recorded once per hour, resulting in 8760 load values ​​(in kilowatts). For example, the load in a certain hour is 50.2 kilowatts. At the same time, external environmental data such as temperature (degrees Celsius) and humidity (percentage) for the corresponding time period are collected. For example, the temperature in a certain hour is 25.3 degrees and the humidity is 60%.

[0040] Adaptive clustering methods were used to group the data. The K-means clustering algorithm was used to extract features from the data according to three dimensions: load, temperature, and humidity. The initial cluster center K was set to 5. The distance from each data point to the center was calculated using Euclidean distance. For example, the distance from a data point to a certain center was 3.5. After multiple iterations and optimizations, the data was finally divided into 5 similar time period feature groups. One group may represent the high load period under high temperature and high humidity, with an average load of about 60.8 kilowatts, an average temperature of 28.5 degrees, and an average humidity of 75%.

[0041] Subsequently, a preliminary fluctuation pattern classification was performed for each group of characteristics, and the load change trend of each group over time was analyzed. For example, by analyzing the time domain characteristics of each load sequence (such as calculating the correlation between the 24-hour moving average and the original sequence), it was found that it exhibits a clear fluctuation pattern with a 24-hour cycle, corresponding to the daily charging peak. Further analysis combined with business scenarios suggests that it may be related to users' commuting times. The load peak occurs around 18:00, with a peak load of 65.3 kW, while the trough occurs at 3:00 AM, with a load of only 20.1 kW. Ultimately, five fluctuation pattern classifications were formed, such as "weekday peak pattern" or "weekend off-peak pattern," providing data support for subsequent load forecasting.

[0042] S102: Based on the preliminary fluctuation pattern classification, the sliding window technique is used to scan continuous data segments in the sequence to determine the short-term sensitivity index in the current period.

[0043] A sliding window technique is used to scan the load sequence segment by segment, extracting short-term load fluctuation amplitude, frequency, and direction indices within each window to obtain a short-term sensitivity sequence. The fluctuation intensity level is determined by comparing the values ​​of each time period in the short-term sensitivity sequence. If the short-term sensitivity sequence value exceeds a preset threshold, it is marked as a high-sensitivity period; otherwise, it is marked as a low-sensitivity period. For load data segments marked as high-sensitivity periods, points where the absolute value of continuous differences exceeds a preset abrupt change threshold are detected, obtaining the location of load abrupt change points and their corresponding abrupt change amplitude values ​​to determine the short-term sensitivity peak interval. Environmental factor sequences are collected from the time range corresponding to the short-term sensitivity peak interval and concatenated with the load sequence within the interval to form a sensitive environment-load joint sequence. Density clustering is used to group all sensitive environment-load joint sequences, obtaining cluster centers and sub-pattern labels for each joint sequence. High-sensitivity periods are categorized based on the sub-pattern labels to determine the short-term sensitive fluctuation sub-pattern for each segment. The short-term sensitive fluctuation sub-patterns are statistically summarized and sorted from high to low frequency to obtain the set of short-term sensitive fluctuation patterns dominated by the current load sequence.

[0044] Based on the preliminary fluctuation pattern classification results, the system employs a sliding window technique to continuously scan and analyze the load sequence under each pattern to quantify the short-term sensitivity index for the current period. First, the window length is set to 168 hours, or one week of data, corresponding to a complete weekday and weekend cycle. For example, for the "weekday peak pattern" sequence, 168 consecutive hours of load data are taken starting from time point t. The standard deviation σ of the load sequence within the window is calculated to be 12.4 kW, and the range of the difference between the maximum and minimum values ​​is calculated to be 48.7 kW. Next, a rate of change index is introduced. The load increment between adjacent hours is calculated using first-order differencing. For example, if the load increases from 42.1 kW to 58.6 kW for two consecutive hours within a window, the increment is 16.5 kW, corresponding to an hourly rate of change of 39.2%. Then, the absolute values ​​of all increments within the window are taken and averaged to obtain an average hourly fluctuation amplitude of 8.7 kW / hour.

[0045] Further calculation of the autocorrelation coefficient of the load series showed that the correlation coefficient ρ1 was 0.82 at a 1-hour lag and ρ1 was 0.82 at a 24-hour lag. 24The value of 0.65 indicates a strong daily cycle correlation. The above indicators were then normalized, mapping the standard deviation, range, and average volatility to the 0-1 interval. For example, the normalized standard deviation was 0.68, the normalized range was 0.75, and the normalized average volatility was 0.61. A weighted fusion method was then used, with weights set to 0.4, 0.3, and 0.3 respectively, to calculate the short-term sensitivity index S = 0.4 × 0.68 + 0.3 × 0.75 + 0.3 × 0.61 = 0.69, indicating that the current load is relatively sensitive to recent disturbances.

[0046] The above calculations are automatically performed for each sliding window and the sensitivity sequence is updated. When the sensitivity exceeds the 0.65 threshold, it is marked as a high-sensitivity period for subsequent dynamic adjustment of the prediction model parameters. The entire process is completed by the background algorithm in a timed loop, realizing real-time quantitative characterization of the short-term fluctuation characteristics of charging load.

[0047] S103: If the short-term sensitivity index exceeds the preset threshold, user behavior data will be integrated through the information processing stage to determine whether there is a risk of sudden change and obtain the adjusted window length range.

[0048] By continuously scanning along the load sequence using a sliding window, the short-term load change rate, short-term load change cycle, and short-term load change trend direction within each window are extracted to obtain a short-term change-sensitive sequence. The values ​​of each segment in the short-term change-sensitive sequence are compared with preset thresholds to determine whether the current segment belongs to a high-change-sensitive period, thus obtaining a set of high-change-sensitive periods. For high-change-sensitive periods, load sequence segments are obtained from the corresponding time range, and user electricity consumption behavior sequences within the same time range are accessed through an information processing step to obtain user electricity consumption behavior sequences for high-change-sensitive periods. If points in the user electricity consumption behavior sequence have consecutive absolute differences exceeding a preset jump threshold, the high-change-sensitive period is marked as having an abnormal jump risk, resulting in a set of high-change-sensitive periods with risk labels. For high-change-sensitive periods with risk labels, the k-means clustering method is used to group the joint sequence obtained by splicing the load sequence segments and user electricity consumption behavior sequences for this period, obtaining cluster centers and risk sub-pattern labels for each joint sequence. Based on the risk sub-pattern labels, the high-change-sensitive periods with risk labels are classified to determine the short-term abnormal change sub-pattern to which each segment belongs. The short-term abnormal change sub-patterns are sorted from high to low frequency to obtain the set of short-term abnormal change patterns dominated by the current load sequence.

[0049] In the field of charging load analysis, when the system detects that the short-term sensitivity index exceeds the preset threshold of 0.7, it automatically triggers the information processing stage, integrates user behavior data through algorithms to assess the risk of sudden changes, and dynamically adjusts the window length range.

[0050] The system extracts user charging behavior data from the database over the past 72 hours, including charging frequency and average power per charge. For example, a certain user group has an average charging frequency of 1.2 times per day and an average charging power (load) of 35.6 kilowatts. Then, the system uses a clustering algorithm to divide user behavior into high-frequency and low-frequency categories. The high-frequency user accounts for 62.3% of the total. The system also analyzes the impact of historical load data on sudden changes and finds that the peak load fluctuation of the high-frequency user group reaches 28.9 kilowatts, which is much higher than the 15.4 kilowatts of the low-frequency group. This indicates that high-frequency users may cause sudden change risks.

[0051] The system uses a decision tree model to quantitatively assess mutation risk. Input features include the proportion of user behavior and the peak fluctuation range. The output risk probability is 0.58. If the probability exceeds 0.5, it is determined that there is a potential mutation risk.

[0052] The system adjusts the window length range based on the risk probability and uses linear interpolation to calculate the adjustment coefficient. For example, the coefficient for a risk probability of 0.58 is 1.15. The original window length is adjusted from 96 hours to 110 hours to cover a wider time range and capture potential fluctuations. At the same time, the upper limit of the window length is set to 120 hours to avoid overexpansion.

[0053] Based on the set of short-term abnormal change patterns dominated by the current load sequence, the system dynamically adjusts the length of the analysis window. For example, for a specific risk pattern that occurs frequently, the original window length is adjusted from 96 hours to 110 hours, thereby obtaining the adjusted window length range, and this related information is stored to form a dynamic feedback mechanism.

[0054] S104: Obtain long-term trend data from the adjusted window length range, regroup historical patterns using an adaptive clustering method, and determine the long-term stability correction coefficient.

[0055] Long-cycle load sequence segments are obtained from the adjusted window length range. These segments are then regrouped using an adaptive clustering method to obtain a set of long-term load patterns. The frequency of occurrence and stability indices of each pattern in the long-term load pattern set are compared to determine the dominant long-term pattern of the current load sequence, resulting in a long-term stability correction coefficient. For each dominant long-term pattern, the periodic characteristic sequence and fluctuation amplitude sequence of the load sequence are extracted from the corresponding time period. The environmental impact sequence of the long-term dominant pattern is obtained by integrating the concurrent external environmental sequence through information processing. If the fluctuation amplitude of consecutive segments in the environmental impact sequence exceeds a preset environmental disturbance threshold, the dominant long-term pattern is marked as having external interference risk, resulting in a set of long-term dominant patterns with interference labels. For these long-term dominant patterns with interference labels, k-means clustering is used to group the joint sequence (consisting of the periodic characteristic sequence, fluctuation amplitude sequence, and environmental impact sequence) to obtain the cluster centers of the interference sub-patterns and the long-term interference sub-type label for each joint sequence.

[0056] In practical implementation, in the field of charging load analysis, after obtaining long-term trend data from the adjusted window length range, the system automatically extracts the charging load sequence of the past 180 days as the basic dataset. For example, the average daily total charging volume gradually increased from the initial 4200 kWh to the recent 5100 kWh, showing a slow growth trend.

[0057] Next, an adaptive clustering method was used to regroup the historical patterns. The system first calculated the shape feature vector of each daily load curve, including the peak time, peak-to-valley ratio, and slope change rate. Then, the optimal number of clusters was automatically determined using a Gaussian mixture model combined with the Bayesian information criterion. The results were divided into three categories: stable, daytime peak, and nighttime concentrated. The stable type accounted for 47.8%, the daytime peak type accounted for 31.4%, and the nighttime concentrated type accounted for 20.8%.

[0058] The system then extracts long-term stability indices for each cluster type. For example, the standard deviation of the load for the stable type is 185 kWh per day, the standard deviation for the daytime peak type is 362 kWh per day, and the standard deviation for the nighttime concentrated type is 418 kWh per day. Based on these indices, the fluctuation coefficients for each cluster type are calculated, and a weighted sum is obtained to obtain an overall long-term stability correction coefficient of 0.87. The system further uses this correction coefficient as an adjustment factor to smooth the real-time load sequence or adjust the parameters of the trend extraction model to reduce over-response to short-term noise while preserving long-term trend characteristics. For example, in the exponential smoothing algorithm, the smoothing factor is multiplied by this correction coefficient (e.g., 0.3 * 0.87 = 0.261) to optimize the preprocessing effect of the sequence.

[0059] Finally, the system stores the obtained correction coefficients and the current cluster distribution ratio into the dynamic parameter table, forming a closed-loop update mechanism. When the next long-cycle data is updated, if the proportion of the concentrated clusters at night rises to 25.6%, the system will re-trigger adaptive clustering and coefficient calculation to ensure that the load analysis always matches the long-term behavioral evolution characteristics of the user group.

[0060] S105: The real-time load sequence is processed by the long-term stability correction coefficient to obtain the optimized multi-timescale prediction model input.

[0061] The real-time load data is initially processed using a long-term stability correction coefficient to obtain adjusted load sequence data. For the adjusted load sequence data, time-series characteristics at multiple time scales are extracted to obtain load change patterns at different time scales. Based on these load change patterns, an input feature set for the multi-time-scale prediction model is constructed, and the feature combination form is determined. The support vector machine algorithm is used to train the input feature set, obtaining the preliminary structure of the multi-time-scale prediction model. If the prediction error of the trained model at a certain time scale exceeds a preset threshold, the feature weights for that time scale are readjusted through information processing to obtain an updated feature combination. The multi-time-scale prediction model is then optimized using the updated feature combination to obtain the final multi-time-scale prediction result. Based on the final multi-time-scale prediction result, corresponding load prediction output data is generated.

[0062] In the field of charging load analysis, the system processes real-time load sequences using long-term stability correction coefficients, and the specific implementation method for optimizing the input of multi-time-scale prediction models is as follows:

[0063] The charging load sequence of the most recent 24 hours is extracted from the real-time data stream. The data sampling frequency is once every 15 minutes, with a total of 96 data points. For example, the load peak on a certain day reaches 6200 kWh and the valley value is as low as 2100 kWh, with a large fluctuation range.

[0064] The pre-calculated long-term stability correction coefficient (e.g., 0.92) was applied to the smoothing of the real-time sequence. The exponential moving average (EMA) algorithm was used to correct the original data. The initial weight was set to 0.4, and after coefficient adjustment, the weight was reduced to 0.368. The smoothed load sequence was calculated, in which the peak value was reduced to 5980 kWh and the valley value was increased to 2280 kWh, reducing short-term noise interference.

[0065] Based on the smoothed sequence, combined with multi-timescale decomposition algorithms (such as wavelet transform), the data is split into feature components at three time scales: hourly (short-term), daily (medium-term), and weekly (long-term). For example, the standard deviation of the hourly component is 245 kWh, the daily component is 380 kWh, and the weekly component is 510 kWh. The analysis shows that short-term fluctuations are mainly driven by random user behavior, while the long-term component reflects the growth trend of basic demand.

[0066] To ensure the data quality of the input prediction model, outlier detection is further performed on each scale component. Outliers are removed using a rule based on three times the standard deviation. For example, two out-of-range values ​​(values ​​exceeding the mean plus or minus three times the standard deviation, i.e., values ​​exceeding 1200 kWh) are removed from the hourly component. Missing data are then filled in using linear interpolation, ultimately generating a smooth and stable multi-scale input dataset.

[0067] The processed data is compared with a historical load pattern database, and the input weights of the prediction model are automatically adjusted. For example, the weight of the short-term component is reduced from 0.5 to 0.45 to enhance the focus on medium-term trends, forming an optimized input matrix for subsequent prediction models. This process is completed through automated algorithms, ensuring the continuity and logic of data processing, while also being closely related to the analysis needs of user behavior changes.

[0068] S106: If the input display deviation of the optimized multi-timescale prediction model increases, the information processing stage is used to analyze seasonal change data, determine the priority of regulation needs, and obtain the final window dynamic configuration scheme.

[0069] A long-term stability correction coefficient is applied to the real-time load sequence to obtain a corrected load sequence. Multi-scale periodic components are extracted from the corrected load sequence to obtain load sequences at each scale. Input feature sets are constructed for each scale load sequence to determine the feature input structure.

[0070] Support vector machines are used to train the feature input structure, obtaining preliminary parameters for the multi-scale prediction model. If the prediction deviation at a certain scale exceeds a set range, seasonal cycle data for that scale is extracted through information processing. Based on the seasonal cycle data, the priority of seasonal regulation is determined, resulting in a window dynamic adjustment configuration scheme. The preliminary parameters of the multi-scale prediction model are then corrected according to the window dynamic adjustment configuration scheme, yielding an optimized multi-scale load prediction sequence.

[0071] In the field of charging load analysis, if the input deviation of the optimized multi-timescale prediction model increases, the system analyzes seasonal variation data through information processing to determine the priority of control demand, and the specific implementation method of the final window dynamic configuration scheme is as follows:

[0072] First, extract the total charging load data for the current month and the two months prior from the historical database. For example, the average daily load in January is 4850 kWh, in December it is 4620 kWh, and in November it is 4410 kWh. Calculate the seasonal growth rate sequence, where the month-on-month growth from December to January is 4.98%, and the month-on-month growth from November to December is 4.76%.

[0073] The Holt-Winters additive seasonal decomposition algorithm was used to process the daily load series over the past 90 days. The smoothing parameter α was set to 0.28, the trend parameter β was set to 0.15, and the seasonal parameter γ was set to 0.62. The seasonal index component was separated. For example, the seasonal index for the current month is 1.074, indicating that the winter electricity demand is 7.4% higher than the baseline level. The standard deviation of the residual component is only 92 kWh, confirming that the seasonal factor is the main source of the deviation.

[0074] Calculate the priority of control demand for each time period and sort them using the load deviation contribution formula (absolute value of deviation multiplied by seasonal index weight). For example, the deviation contribution of the evening peak period from 18:00 to 20:00 is 0.037, so it has the highest priority, while the contribution of the midday trough period from 12:00 to 14:00 is 0.009, so it has the lowest priority.

[0075] To achieve precise regulation, the system dynamically adjusts the prediction window length based on priority, employing an adaptive sliding window mechanism. High-priority periods have their windows shortened to 4 hours (originally 8 hours), while low-priority periods are extended to 12 hours. Simultaneously, a secondary smoothing weight of 0.55 is added to the data within the high-priority window, creating a tiered regulation intensity configuration scheme. For example, a smoothing coefficient of 0.55 is applied within the evening peak window, while the midday window maintains a weight of 0.38. Finally, a dynamic configuration table containing the window length and smoothing coefficient for each period is generated for the prediction model to access in real time. This entire process is executed automatically by the background algorithm, ensuring timely response to seasonal changes and effective matching with grid regulation needs.

[0076] S107: Based on the final window dynamic configuration scheme, obtain the power grid operation status feedback, iteratively update the load balance parameters through sliding window technology, and determine the basis for resource efficiency allocation.

[0077] Acquire power grid operation status feedback data. Extract periodic fluctuation data through information processing to obtain a periodic fluctuation sequence. Determine the priority order of resource regulation based on the periodic fluctuation sequence and identify priority regulation targets. If the real-time status of resource allocation corresponding to the priority regulation target shows a deviation, iteratively update the load balance parameters using a sliding window technique to obtain updated load balance parameters. Determine the basis for resource efficiency allocation based on the updated load balance parameters and obtain a dynamic resource allocation scheme. Implement targeted corrections to the load balance parameters based on the dynamic resource allocation scheme to obtain an optimized load balance sequence. Extract the current resource allocation status based on the optimized load balance sequence and determine whether it meets the equilibrium conditions. If the current resource allocation status does not meet the equilibrium conditions, return to the information processing stage to re-extract periodic fluctuation data. If the current resource allocation status meets the equilibrium conditions, output the final resource efficiency allocation basis.

[0078] In the field of charging load analysis, based on the final window dynamic configuration scheme, the system obtains real-time feedback data on the grid operation status, including the current bus voltage deviation, transformer load rate, and the fluctuation value of new energy output in the region. For example, the voltage deviation is -3.2%, the transformer load rate is 87.6%, and the photovoltaic output is 12.4% lower than predicted. The load balance parameters of the most recent 48 hours are iteratively updated through sliding window technology.

[0079] First, the grid feedback sequence is extracted, and the load balance deviation is processed using an exponentially weighted moving average algorithm with a smoothing coefficient λ set to 0.32. The current balance offset is calculated to be -185 kWh. Then, a sliding window is constructed with a 5-minute step size. The difference between the actual load and the predicted load for each time period within the window is accumulated. The balance parameter is then adjusted in conjunction with the fluctuation rate of renewable energy output. For example, the accumulated deviation in the evening peak window reaches -420 kWh, and the adjusted balance parameter is 0.78, while the accumulated deviation in the midday off-peak window is only -65 kWh, and the adjusted balance parameter is 0.94.

[0080] The calculation of resource efficiency allocation is based on the efficiency priority ranking formula (balance parameter × load density × new energy matching degree) to obtain the scheduling priority of each resource. For example, the efficiency score of energy storage resources is 0.652 during the evening peak period, with the highest priority, while the efficiency score of distributed photovoltaics is 0.218 during midday, with the lowest priority.

[0081] Based on priority, a resource allocation ratio table is automatically generated. During high-priority periods, the proportion of energy storage discharge power is increased to 65%, while during low-priority periods it remains at 22%. The updated balance parameters and allocation ratios are written into the model input layer in real time to achieve dynamic matching of load and resources. The entire iterative update process is automatically triggered by the background control module every 15 minutes.

[0082] S108: Obtain peak load data from resource efficiency allocation criteria, use adaptive clustering to classify the risk of waste during off-peak periods, and obtain the demand regulation sequence under green transformation.

[0083] Peak-hour load data is obtained from resource efficiency allocation criteria. Off-peak load data is classified using an adaptive clustering method to identify off-peak waste risk categories. Demand control sequences are then defined based on these categories. Off-peak load distribution characteristics are extracted from the demand control sequences. Priority control targets are determined based on these characteristics. If the off-peak load corresponding to a priority control target shows a waste risk exceeding a preset range, time-series smoothing techniques are used to process the demand control sequences, resulting in a corrected sequence. The corrected demand control sequence is used to determine whether the current off-peak load control status has reached the green equilibrium condition. If the green equilibrium condition is met, the final green transformation control criteria are output.

[0084] In the field of power grid load management, the system extracts peak load data from the resource efficiency allocation criteria. For the load records of a certain area during the evening peak period, it obtains the peak electricity consumption data from 17:00 to 21:00. For example, if the total load is 5200 kilowatts and the peak load density reaches 1.35 kilowatts / square kilometer, the data is automatically imported into the analysis module. Through the time series decomposition algorithm, the load data is split into trend components and periodic components, and the periodic peak fluctuation rate is calculated to be 8.7%, providing a data basis for subsequent classification.

[0085] Next, an adaptive clustering method was used to classify the waste risk during off-peak hours. Based on the K-means clustering algorithm, load utilization rate and renewable energy absorption rate were used as feature dimensions. The initial cluster centers were set to 3. After iterative calculation, the off-peak hours were divided into three categories: high risk, medium risk and low risk. For example, the load utilization rate from 2:00 to 5:00 am was only 21.3%, and the renewable energy absorption rate was as low as 15.6%, which was classified as a high-risk waste category. On the other hand, the utilization rate from 6:00 to 8:00 am was 45.2%, and the absorption rate was 38.9%, which was classified as a medium-risk category.

[0086] Further combining the classification results, the impact of off-peak waste risk on overall power grid efficiency is analyzed, and a risk weight matrix is ​​automatically generated. For example, the weight of high-risk periods is set to 0.65, and medium-risk periods are set to 0.35. The comprehensive waste risk index is calculated to be 0.48 through weighted calculation.

[0087] Based on the goal of green transformation, a demand regulation sequence is constructed, and the load transfer scheme is optimized using a linear programming algorithm. The constraint is set that the renewable energy consumption rate is not less than 50%. The calculation shows that the load to be transferred during high-risk periods is 320 kW, and the load to be transferred during low-risk periods is 80 kW. The sequence data is automatically stored in the regulation database and linked with the regional energy conservation and emission reduction targets to ensure that the power grid operation meets the needs of green development. The entire process is updated by the system backend every 4 hours, forming a closed-loop management.

[0088] This invention brings significant benefits to the fields of electric vehicle charging load forecasting and power grid dynamic balance by integrating multi-timescale adaptive window adjustment and long-term and short-term stability collaborative correction mechanisms.

[0089] This invention utilizes an adaptive clustering method to fuse features and identify patterns from historical loads and external environmental factors, constructing a precise preliminary classification foundation for fluctuation patterns. Furthermore, it employs a sliding window technique to dynamically scan load sequences, quantifying short-term sensitivity indicators in real time. When these indicators exceed thresholds, it integrates user behavior data to assess the risk of sudden changes, achieving keen perception and early warning of sudden load fluctuations. By introducing a long-term stability correction coefficient to optimize real-time sequences, it effectively removes short-term noise interference and enhances long-term trend characteristics, fundamentally improving the input quality of multi-timescale prediction models and significantly reducing prediction bias at different time scales.

[0090] Meanwhile, this invention determines the priority of control needs by analyzing seasonal variation data and generates a final dynamic window configuration scheme accordingly, enabling the prediction model to adaptively match macro-cyclical patterns. Furthermore, based on real-time power grid operation feedback, iterative updates to load balance parameters are achieved using sliding window technology, realizing dynamic self-correction of the scheduling strategy and ensuring the timeliness and accuracy of resource allocation. By identifying peak load characteristics and clustering waste risks during off-peak periods, a clearly defined green transformation demand control sequence can be generated, thereby effectively promoting peak shaving and valley filling and the efficient consumption of new energy sources while ensuring the safe and stable operation of the power grid.

[0091] In summary, this invention systematically integrates short-term fluctuation sensitivity analysis, long-term trend stability correction, and multi-scale window dynamic configuration to form an adaptive, high-precision, and closed-loop optimizable charging load management method. This method not only significantly improves the reliability of load forecasting and the foresight of grid dispatching, but also effectively optimizes the allocation efficiency of power resources in the spatiotemporal dimensions. It provides core technical support for the grid to cope with the challenges of large-scale electric vehicle integration and makes a substantial contribution to promoting the green and low-carbon transformation of the energy system.

[0092] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0093] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic charging load balancing method based on a time-series prediction algorithm, characterized in that, The method includes: S101: Obtain real-time load sequences and external environmental factors from historical charging data, and group similar time period features using an adaptive clustering method to obtain a preliminary fluctuation pattern classification; S102: Based on the preliminary fluctuation pattern classification, the sliding window technique is used to scan continuous data segments in the sequence to determine the short-term sensitivity index in the current period. S103: If the short-term sensitivity index exceeds the preset threshold, user behavior data will be integrated through the information processing stage to determine whether there is a risk of sudden change and to obtain the adjusted window length range. S104: Obtain long-term trend data from the adjusted window length range, regroup historical patterns using an adaptive clustering method, and determine the long-term stability correction coefficient; S105: The real-time load sequence is processed by the long-term stability correction coefficient to obtain the optimized multi-time-scale prediction model input; S106: If the input display deviation of the optimized multi-timescale prediction model increases, the information processing stage is used to analyze seasonal change data, determine the priority of regulation needs, and obtain the final window dynamic configuration scheme. S107: Based on the final window dynamic configuration scheme, obtain the power grid operation status feedback, iteratively update the load balance parameters through sliding window technology, and determine the basis for resource efficiency allocation; S108: Obtain peak load data from resource efficiency allocation criteria, use adaptive clustering to classify the risk of waste during off-peak periods, and obtain the demand regulation sequence under green transformation.

2. The charging load dynamic balancing method based on time-series prediction algorithm according to claim 1, characterized in that, The process of obtaining real-time load sequences and external environmental factors from historical charging data, and grouping similar time-period features using an adaptive clustering method to obtain a preliminary fluctuation pattern classification includes: Obtain real-time load sequences and external environmental factors from historical charging data; The real-time load sequence is divided into multiple continuous time period sequences using a time-series segmentation method. For each continuous time period sequence, the average load, peak load, and rate of change of load are extracted as load feature vectors. Collect the numerical values ​​of external environmental factors corresponding to each consecutive time period sequence to form an environmental feature vector; The load feature vector and the environmental feature vector are concatenated to obtain the comprehensive feature vector for each time period; An adaptive clustering method is used to group all the comprehensive feature vectors to obtain the cluster centers and the cluster labels for each time period; All time-period sequences are categorized based on clustering labels to determine preliminary fluctuation pattern classifications.

3. The charging load dynamic balancing method based on time-series prediction algorithm according to claim 1, characterized in that, The process of classifying preliminary fluctuation patterns and using a sliding window technique to scan continuous data segments in the sequence to determine short-term sensitivity indicators for the current period includes: The sliding window technique is used to scan along the load sequence segment by segment, and the short-term load fluctuation amplitude, short-term load fluctuation frequency and short-term load fluctuation direction index are extracted in each window to obtain the short-term sensitivity sequence. The fluctuation intensity level is determined by comparing the values ​​of each period in the short-term sensitivity sequence. If the value of the short-term sensitivity sequence exceeds a preset threshold, it is marked as a high-sensitivity period; otherwise, it is marked as a low-sensitivity period. For load data segments marked as high-sensitivity periods, detect points where the absolute value of continuous difference within the sequence exceeds a preset mutation threshold, obtain the location of load mutation points and the corresponding mutation amplitude value, and determine the short-term sensitivity peak range. Environmental factor sequences are collected from the time range corresponding to the short-term sensitivity peak interval, and spliced ​​with the load sequence within the interval to form a sensitive environment-load joint sequence; Density clustering was used to group all sensitive environment-load joint sequences, and cluster centers and sub-pattern labels of each joint sequence were obtained. Highly sensitive periods are categorized based on sub-pattern labels to determine the short-term sensitive fluctuation sub-pattern to which each period belongs; By statistically summarizing the short-term sensitive fluctuation sub-patterns and sorting them by frequency of occurrence from high to low, a set of short-term sensitive fluctuation patterns dominated by the current load sequence is obtained.

4. The charging load dynamic balancing method based on time-series prediction algorithm according to claim 1, characterized in that, If the short-term sensitivity index exceeds a preset threshold, user behavior data is integrated through information processing to determine whether there is a risk of sudden change, and an adjusted window length range is obtained, including: By continuously scanning along the load sequence through a sliding window, the short-term load change rate, short-term load change cycle, and short-term load change trend direction within each window are extracted to obtain a short-term change sensitive sequence. By comparing the values ​​of each segment in the short-term change sensitive sequence with a preset threshold, it is determined whether the current segment belongs to a high change sensitive period, and a set of high change sensitive periods is obtained. For periods of high sensitivity to change, load sequence segments are obtained from the corresponding time range, and user electricity consumption behavior sequences within the same time range are accessed through the information processing stage to obtain user electricity consumption behavior sequences for periods of high sensitivity to change. If a point in the user's electricity consumption behavior sequence has a continuous difference absolute value that exceeds the preset jump threshold, then the high change sensitive period is marked as having an abnormal jump risk, and a set of high change sensitive periods with risk labels is obtained. For high-change-sensitive periods with risk labels, the k-means clustering method is used to group the joint sequence of load sequence fragments and user electricity consumption behavior sequences for this period, and obtain the cluster centers and the risk sub-pattern labels to which each joint sequence belongs; Based on the risk sub-pattern labels, the high-change-sensitive periods marked with risk are classified to determine the short-term abnormal change sub-pattern to which each period belongs; The short-term abnormal change sub-patterns are sorted from high to low frequency to obtain the set of short-term abnormal change patterns dominated by the current load sequence.

5. The charging load dynamic balancing method based on time-series prediction algorithm according to claim 1, characterized in that, The process of obtaining long-term trend data from the adjusted window length range, regrouping historical patterns using an adaptive clustering method, and determining long-term stability correction coefficients includes: Obtain long-cycle load sequence segments from the adjusted window length range; Long-term load sequence fragments are regrouped using an adaptive clustering method to obtain a set of long-term load patterns; By comparing the frequency of occurrence of each pattern in the long-term load pattern set with stability indices, the long-term dominant pattern to which the current load sequence belongs is determined, and the long-term stability correction coefficient is obtained. For the long-term dominant pattern, the periodic feature sequence and fluctuation amplitude sequence of the load sequence are extracted from the corresponding time period; By accessing the concurrent external environment sequence through the information processing stage, the environmental impact sequence of the long-term dominant pattern can be obtained. If the fluctuation amplitude of a continuous segment in the environmental impact sequence exceeds the preset environmental disturbance threshold, the long-term dominant mode is marked as having an external interference risk, resulting in a set of long-term dominant modes with interference labels. For long-term dominant patterns with interference labels, the k-means clustering method is used to group the joint sequences of periodic feature sequences, fluctuation amplitude sequences and environmental influence sequences, and obtain the cluster centers of interference sub-patterns and the long-term interference sub-type labels of each joint sequence.

6. The charging load dynamic balancing method based on time-series prediction algorithm according to claim 1, characterized in that, The process of processing the real-time load sequence using a long-term stability correction coefficient to obtain the optimized multi-timescale prediction model input includes: The real-time load data is preliminarily processed using a long-term stability correction coefficient to obtain adjusted load sequence data; For the adjusted load sequence data, time-series characteristics at multiple time scales are extracted to obtain load change patterns at different time scales; Based on load change patterns at different time scales, a set of input features for a multi-time-scale prediction model is constructed, and the combination of features is determined. The support vector machine algorithm is used to train the input feature set to obtain the preliminary structure of the multi-time scale prediction model; If the prediction error of the trained model exceeds a preset threshold at a certain time scale, the feature weights at that time scale will be readjusted through the information processing stage to obtain an updated feature combination. The multi-timescale prediction model is optimized by updating the feature combination to obtain the final multi-timescale prediction result. Based on the final multi-timescale forecast results, corresponding load forecast output data is generated.

7. The charging load dynamic balancing method based on time-series prediction algorithm according to claim 1, characterized in that, If the input deviation of the optimized multi-timescale prediction model increases, the information processing stage is used to analyze seasonal variation data, determine the priority of regulation needs, and obtain the final dynamic window configuration scheme, including: The real-time load sequence is processed using a long-term stability correction factor to obtain a corrected load sequence; Multi-scale periodic components are extracted from the modified load sequence to obtain load sequences at each scale; For load sequences at various scales, input feature groups are constructed to determine the feature input structure; The feature input structure is trained using a support vector machine to obtain the preliminary parameters of the multi-scale prediction model; If the prediction deviation at a certain scale exceeds the set range, seasonal cycle data at that scale will be extracted through the information processing stage. Based on seasonal cycle data, determine the priority of seasonal regulation and obtain a dynamic window adjustment configuration scheme; The initial parameters of the multi-scale prediction model are corrected based on the window dynamic adjustment configuration scheme to obtain the optimized multi-scale load prediction sequence.

8. The charging load dynamic balancing method based on time-series prediction algorithm according to claim 1, characterized in that, The process of obtaining power grid operation status feedback based on the final window dynamic configuration scheme, iteratively updating load balancing parameters through sliding window technology, and determining the basis for resource efficiency allocation includes: Obtain power grid operation status feedback data; Periodic fluctuation data are extracted through information processing to obtain periodic fluctuation sequences; Determine the priority order of resource regulation based on the periodic fluctuation sequence, and identify the priority regulation targets; If the real-time status of resource allocation for the priority control object shows a deviation, the load balancing parameters are iteratively updated using the sliding window technique to obtain the updated load balancing parameters. Based on the updated load balancing parameters, determine the basis for resource efficiency allocation and obtain a dynamic resource configuration scheme; Based on the dynamic resource allocation scheme, the load balancing parameters are adjusted accordingly to obtain the optimized load balancing sequence; Extract the current resource allocation status based on the optimized load balancing sequence and determine whether the balance condition is met. If the current resource allocation status does not meet the equilibrium condition, return to the information processing stage to re-extract the periodic fluctuation data; If the current resource allocation status meets the equilibrium condition, then the final resource efficiency allocation basis will be output.

9. The charging load dynamic balancing method based on time-series prediction algorithm according to claim 1, characterized in that, The process involves obtaining peak-hour load data from resource efficiency allocation criteria, classifying off-peak waste risks using an adaptive clustering method, and deriving a demand regulation sequence under green transformation, including: Obtain peak load data from resource efficiency allocation criteria; The load data during off-peak hours is classified using an adaptive clustering method to obtain off-peak waste risk categories; Demand control sequences are categorized according to the risk type of waste during off-peak periods; Extract load distribution characteristics during off-peak periods from the demand regulation sequence; Determine the priority targets for regulation based on load distribution characteristics; If the waste risk of the off-peak load corresponding to the control priority object exceeds the preset range, time series smoothing technology is used to process the demand control sequence to obtain a corrected demand control sequence. Based on the revised demand control sequence, determine whether the current off-peak load control status has reached the green equilibrium condition. If the green equilibrium condition has been reached, output the final green transformation control basis.

10. The charging load dynamic balancing method based on time-series prediction algorithm according to any one of claims 1-9, characterized in that, The following system will be used to implement this: The system includes: The data acquisition and pattern classification module is used to obtain real-time load sequences and external environmental factors from historical charging data, and to obtain preliminary fluctuation pattern classification by grouping similar time period features through an adaptive clustering method. The short-term sensitivity analysis module, connected to the data acquisition and pattern classification module, is used to determine the short-term sensitivity index in the current period by scanning continuous data segments in the sequence using sliding window technology based on the preliminary fluctuation pattern classification. The dynamic window adjustment module, connected to the short-term sensitivity analysis module, is used to integrate user behavior data to determine whether there is a risk of sudden change when the short-term sensitivity index exceeds a preset threshold, and generate an adjusted window length range accordingly. The long-term stability correction module, connected to the dynamic window adjustment module, is used to obtain long-term trend data from the adjusted window length range, regroup historical patterns using an adaptive clustering method, and determine the long-term stability correction coefficient. A multi-timescale prediction input optimization module is connected to the long-term stability correction module and is used to process the real-time load sequence through the long-term stability correction coefficient to obtain the optimized multi-timescale prediction model input. The regulation strategy generation module is connected to the multi-timescale prediction input optimization module. It is used to analyze seasonal change data to determine the priority of regulation needs and generate the final window dynamic configuration scheme when the input display deviation of the optimized multi-timescale prediction model increases. The load parameter iteration and resource allocation module is connected to the control strategy generation module. It is used to obtain power grid operation status feedback according to the final window dynamic configuration scheme, and to iteratively update the load balance parameters through sliding window technology to determine the basis for resource efficiency allocation. The green regulation sequence generation module is connected to the load parameter iteration and resource allocation module. It is used to obtain peak load data from the resource efficiency allocation basis and use an adaptive clustering method to classify the waste risk during off-peak periods to obtain the demand regulation sequence under green transformation.

Images