Time-series decomposition-based charging load prediction method and device, and electronic equipment

Abstract

The application discloses a kind of based on timing decomposition's charging load prediction method, device and electronic equipment, belong to load prediction technical field, to solve the problem that traditional prediction method is difficult to effectively capture the complex seasonal rule of electric vehicle charging. Method includes: obtaining historical load data, historical load data includes load value and meteorological environment information in preset period;Prior characteristics and basic characteristics are obtained based on historical load data, basic characteristics include at least one of time index feature and meteorological environment feature, prior characteristics are obtained by decomposing the time series characteristics of historical load data by Prophet model;Prior characteristics and basic characteristics are fused to obtain input feature vector;Input feature vector is input into pre-trained prediction module to carry out load prediction, and prediction result is obtained.The application embodiment can improve the accuracy of charging load prediction result.

Description

Technical Field

[0001] This invention belongs to the field of load forecasting technology, specifically relating to a charging load forecasting method, device, and electronic equipment based on time-series decomposition. Background Technology

[0002] The rapid popularization of new energy vehicles has brought about new growth points in electricity load. This growth has driven a rapid increase in electric vehicle charging load. This large-scale access has led to a high degree of uncertainty and coupling in the charging load in both time and space dimensions. From a microscopic perspective, the charging demand of electric vehicles is not only constrained by users' travel patterns (such as commuting tides and holiday travel), but also exhibits a complex nonlinear relationship with the meteorological environment. For example, extreme high and low temperatures can significantly change the charging and discharging characteristics of power batteries and the energy consumption of in-vehicle air conditioning, thus distorting the load curve. In addition, different types of charging stations (such as slow charging in residential areas and fast charging stations on highways) have drastically different daily load patterns. This intertwined correlation and heterogeneity of multiple factors means that the charging load sequence contains deep-seated periodic patterns and abrupt changes that are difficult to capture by traditional linear models.

[0003] Constructing effective feature variables and uncovering the inherent patterns of electric vehicle charging load is key to improving prediction accuracy. Existing prediction methods struggle to effectively capture the complex seasonal patterns of electric vehicle charging, resulting in low accuracy in charging load predictions. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that the methods commonly used in the prior art are difficult to effectively capture the complex seasonal patterns of electric vehicle charging, resulting in low accuracy of charging load prediction results. In order to solve the above problems, the present invention provides a charging load prediction method, device and electronic device based on time-series decomposition.

[0005] The content of this invention includes:

[0006] In a first aspect, embodiments of the present invention provide a charging load prediction method based on time-series decomposition, comprising:

[0007] Acquire historical load data, which includes load values ​​and meteorological information within a preset period;

[0008] Based on the historical load data, prior features and basic features are obtained. The basic features include time index features and meteorological environment features. The prior features are obtained by decomposing the time series characteristics of the historical load data using the Prophet model. The time index features are determined based on time information within the preset period. The meteorological environment features are determined based on the meteorological environment information.

[0009] The prior features and the basic features are fused to obtain the input feature vector;

[0010] The input feature vector is input into a pre-trained prediction module to perform load prediction and obtain the prediction result.

[0011] Optionally, obtaining prior features based on the historical load data includes:

[0012] The load value is decomposed into a combination of trend, seasonality, holiday, and error terms using the Prophet model. The trend term is used to fit the non-periodic baseline drift, the seasonality term is used to fit the periodic fluctuations of the load value on the weekly and yearly scales, the holiday term is used to capture load abrupt changes during statutory holidays, and the error term is a random error that follows a normal distribution.

[0013] The value of the trend item at each time point is extracted as a trend component, the value of the seasonality item at each time point is extracted as a seasonal component, the value of the holiday item at each time point is extracted as a holiday component, and a predicted value component is generated based on the trend item, the seasonality item, and the holiday item.

[0014] Determine prior features, which include at least one of the trend component, the seasonal component, the holiday component, and the predicted value component.

[0015] Optionally, the seasonal component includes a daily periodic feature component and a weekly periodic feature component, and the extraction of the seasonal term's value at each time point as the seasonal component includes:

[0016] The seasonality term is decomposed into multi-scale components to extract the daily periodic feature components and the weekly periodic feature components, respectively.

[0017] The daily periodic characteristic component is based on the load waveform of morning and evening peaks and nighttime troughs fitted by a 24-hour Fourier series; the weekly periodic characteristic component is based on the load difference between weekdays and weekends fitted by a 7-day Fourier series.

[0018] Optionally, the time step can be determined using the Prophet model. load value The breakdown into trend, seasonality, holiday, and error components is as follows:

[0019] ;

[0020] Alternatively, the time step can be determined using the Prophet model. load value The breakdown into trend, seasonality, holiday, and error components is as follows:

[0021] ;

[0022] in, For the trend term, For the aforementioned seasonal item, For the aforementioned holiday items, This refers to the error term.

[0023] Optionally, the prediction module is built based on the XGBoost model.

[0024] Optionally, the basic features also include at least one of the following:

[0025] Short-term lag characteristics, which include load values ​​from the 1st to the Nth hour prior to the current time, are used to characterize the short-term autocorrelation of the load, where N is a positive integer;

[0026] The periodic lag feature includes the load values ​​at the M previous periods that have the same periodic phase as the current time, which are used to capture the cyclic dependence of the load on a daily or weekly scale, where M is a positive integer.

[0027] The rolling statistical features are calculated based on multiple preset time windows. The rolling statistical features include at least one of the rolling mean, rolling maximum value and rolling standard deviation of the load sequence within the preset time window, which are used to characterize the central tendency and dispersion of the load within a local time range.

[0028] Optionally, the meteorological environmental information includes temperature, humidity, rainfall, and surface horizontal radiation, and the meteorological environmental characteristics include at least one of the following:

[0029] Basic meteorological characteristics, including temperature characteristics corresponding to the temperature, humidity characteristics corresponding to the humidity, rainfall characteristics corresponding to the precipitation, and radiation characteristics corresponding to the horizontal radiation of the land surface;

[0030] The first Boolean feature is used to characterize whether it is hot weather;

[0031] The second Boolean feature is used to characterize whether or not it has rained;

[0032] The second-order interactive features are obtained by performing pairwise dot products on the normalized basic meteorological features.

[0033] Secondly, embodiments of the present invention provide a charging load prediction device based on time-series decomposition, comprising:

[0034] The first acquisition module is used to acquire historical load data, which includes load values ​​and meteorological information within a preset period.

[0035] The second acquisition module is used to acquire prior features and basic features based on the historical load data. The basic features include time index features and meteorological environment features. The prior features are obtained by decomposing the time series characteristics of the historical load data using the Prophet model. The time index features are determined based on time information within the preset period. The meteorological environment features are determined based on the meteorological environment information.

[0036] The fusion module is used to fuse the prior features and the basic features to obtain the input feature vector;

[0037] The input module is used to input the input feature vector into the pre-trained prediction module for load prediction and obtain the prediction result.

[0038] Thirdly, embodiments of the present invention provide an electronic device, including: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is configured to read the program in the memory to implement the steps in the charging load prediction method based on time-series decomposition as described in the first aspect.

[0039] Fourthly, embodiments of the present invention provide a readable storage medium for storing a program, which, when executed by a processor, implements the steps in the charging load prediction method based on time-series decomposition as described in the first aspect.

[0040] The beneficial effects of this invention are as follows: In this embodiment, the Prophet model is used as a high-order feature extractor to decompose the time-series characteristics of historical load data to obtain prior features. These prior features, together with time index features and meteorological environment features, constitute the input feature vector, providing a strongly constrained time pattern benchmark for the subsequent prediction module. On the one hand, by pre-extracting the main temporal structure through the Prophet model, the prediction module only needs to focus on the residual signal between the actual load and the benchmark provided by the prior features, significantly simplifying the complexity of the objective function, accelerating convergence, and reducing dependence on large-scale training data. On the other hand, the prior features extracted by the Prophet model can effectively guide the prediction module to maintain reasonable estimates at load abrupt change points (such as the first day of resumption of work after a holiday, or extreme weather over a weekend), avoiding drastic biases caused by sample sparsity in purely data-driven models. By integrating the structured time modeling capabilities of the Prophet model with the nonlinear fitting and feature interaction capabilities of the prediction model, a hybrid paradigm of prior guidance and data-driven approaches is formed. This approach retains the stability of statistical models in predicting long-term patterns while leveraging the sensitivity of machine learning to local disturbances (such as sudden temperature changes or temporary promotional activities), thereby significantly improving the accuracy of prediction results. Attached Figure Description

[0041] Figure 1 A flowchart illustrating the charging load prediction method based on time-series decomposition provided in an embodiment of the present invention;

[0042] Figure 2 A comparison chart of the effects of the partitioning strategies provided in the embodiments of the present invention;

[0043] Figure 3 A schematic diagram of the charging load prediction device based on time-series decomposition provided in an embodiment of the present invention;

[0044] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0045] In the embodiments of this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar. The terms "first," "second," etc., in the specification of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such terms can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are usually of the same class, without limiting the number of objects. For example, the first object can be one or multiple.

[0046] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0048] The following is an explanation of some of the terms used in this application.

[0049] Extreme Gradient Boosting (XGBoost) is a highly efficient ensemble learning algorithm based on Gradient Boosting Decision Tree (GBDT). Its core idea is to train multiple weak learners (i.e., Classification and Regression Trees (CART)) through sequential iteration. In each iteration, the objective function of XGBoost is determined by the loss function. and regularization term It consists of two parts. Used to measure the deviation between predicted and actual values. The regularization term controls model complexity to prevent overfitting. XGBoost performs a second-order Taylor expansion of the objective function, using first and second derivative information to guide tree generation, achieving faster convergence and higher accuracy. Furthermore, the model's built-in column sampling and automatic missing value handling mechanism demonstrates excellent robustness when processing the high-dimensional sparse feature set constructed in this paper.

[0050] Random Forest (RF), a typical example of the bagging ensemble strategy, constructs multiple decorrelational decision trees and takes the average as the final output. This model can effectively reduce the variance of a single decision tree, has good adaptability to high-dimensional data, and is not prone to overfitting. It is often used as a strong benchmark model for power load forecasting.

[0051] Long Short-Term Memory (LSTM) networks are a variant of Recurrent Neural Networks (RNNs) specifically designed for processing time-series data. By introducing forget gates, input gates, and output gates, LSTMs can effectively solve the vanishing and exploding gradient problems in long sequence training. They excel at capturing long-distance temporal dependencies in charging load sequences and are currently the mainstream method in deep learning for load prediction.

[0052] The Transformer model, based on self-attention, abandons the recurrent structure of RNNs, enabling parallel processing of sequential data and direct capture of global dependencies. Although the Transformer performs exceptionally well in natural language processing, it often struggles to extract local features when handling short-cycle, high-noise numerical load sequences.

[0053] Based on this, embodiments of this application provide a charging load prediction method, apparatus, and electronic device based on time-series decomposition, aiming to improve the accuracy of charging load prediction. Please refer to... Figure 1 , Figure 1 This is a flowchart illustrating the charging load prediction method based on time-series decomposition provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this method specifically includes the following steps:

[0054] Step 101: Obtain historical load data, which includes load values ​​and meteorological information within a preset period.

[0055] Step 102: Obtain prior features and basic features based on the historical load data. The basic features include time index features and meteorological environment features. The prior features are obtained by decomposing the time series characteristics of the historical load data using the Prophet model. The time index features are determined based on time information within the preset period. The meteorological environment features are determined based on the meteorological environment information.

[0056] Step 103: Fuse the prior features and the basic features to obtain the input feature vector.

[0057] Step 104: Input the input feature vector into the pre-trained prediction module to perform load prediction and obtain the prediction result.

[0058] In some embodiments, historical load data is obtained by collecting multi-source observation records within a continuous time window (i.e., a preset period) through a power dispatching system, a charging facility operation platform, a smart meter database, or an energy management system.

[0059] It should be understood that the length of the preset period can be set and adjusted according to actual needs, and is not limited here. For example, the preset period can be 10 days, 6 months, 1 year, or 2 years. In some embodiments, the length of the preset period must meet two basic requirements. First, the time span must be sufficient to cover typical cyclical patterns, including at least several complete weekly cycles (for learning weekday / weekend differences) and seasonal variations (if an annual cycle is enabled). Second, the data must have good continuity, a low proportion of missing values, or have been preprocessed using methods such as interpolation and smoothing to ensure the integrity of the time series.

[0060] Historical load data includes load values, which refer to the charging load values ​​recorded at a fixed time granularity (usually hourly, such as every hour) within a preset period. The unit can be kilowatt (kW), megawatt (MW), or kilowatt-hour (kWh). Load values ​​directly reflect the spatiotemporal distribution characteristics of user charging behavior.

[0061] Meteorological environmental information refers to external environmental parameters collected synchronously with the load observation time, which mainly include, but are not limited to: temperature, humidity, wind speed, precipitation, sunshine duration or weather type (such as sunny, rainy, snowy), etc.

[0062] It should be understood that meteorological environmental information typically originates from public meteorological service interfaces, third-party application programming interfaces (APIs), or local weather stations, and can also be monitored and collected through corresponding sensors; the specific source is not limited here. In this embodiment, the collection of meteorological environmental information must be strictly aligned with the timestamp of the load value.

[0063] Temperature, humidity, rainfall, and solar radiation are important factors affecting the driving range of electric vehicles and users' willingness to travel. In some embodiments, a raw observation sequence of meteorological environmental information with time granularity consistent with the load value is obtained through a meteorological monitoring interface, mainly including temperature, humidity, rainfall, and surface horizontal radiation. Then, the meteorological environmental information is linked with the load value through a timestamp index to ensure that the load at each moment corresponds to the meteorological environmental state at that time.

[0064] As a specific implementation, the historical load data is organized into a structured time-series dataset, with each row corresponding to a time point. It includes the following fields: Timestamp, Load Value Meteorological and environmental information, as well as other auxiliary information (such as electricity prices, holiday information, etc.).

[0065] In this embodiment, the basic features include time index features, which do not rely on external sensors and are designed to extract structured time-series patterns from unstructured timestamps. Optionally, in some embodiments, obtaining time index features based on the historical load data includes:

[0066] The original time field in the historical load data is parsed and formatted to generate a standard timestamp object, and numerical time attributes are extracted from it. The time attributes include hour, day, month and day of the week.

[0067] Based on the predefined conversion rules between the values ​​of time attributes and binary feature values, the standard time object is converted into a time index feature.

[0068] First, the time index in the original time data is parsed and formatted to a datetime format, resulting in a standard time object. Based on this standard time object, numeric time attributes, such as hour (0-23), date, month (1-12), and day of the week (0-6), are directly parsed out.

[0069] In addition to the time attribute, a binary feature is constructed based on business logic as the time index feature. Optionally, in some embodiments, the time index feature includes a first indicator feature, a second indicator feature, a third indicator feature, and a fourth indicator feature, wherein the first indicator feature is used to describe whether it is a weekend ( The second indicator feature is used to describe whether it is a holiday ( The third indicator feature is used to describe whether it is a peak period ( The fourth indicator feature is used to describe whether it is a trough period ( ).

[0070] Specifically, the corresponding time index feature is determined by checking whether the value of the "weekday" attribute in the standard time object is greater than or equal to 5. The value of "" is determined by matching the date with a pre-defined list of holidays. The value of "" is determined. Simultaneously, time periods are categorized based on the number of hours, such as "morning / evening peak" and "nighttime off-peak," to determine the time index characteristics. "and" The value of "".

[0071] As a specific implementation, the time index features are all obtained through Boolean feature mapping, using binary features of 0 or 1 to indicate different value types. For example, " The value of 0 is used to indicate that it is not the weekend. The value 1 of “” is used to indicate that it is the weekend.

[0072] In some embodiments, for cyclic variables such as hours, days of the week, and months, trigonometric functions are used for encoding to eliminate numerical hard truncation (e.g., 23:00 and 0:00 are numerically discontinuous but physically continuous). Specifically, the time variable is projected onto a unit circle, and its sine (Sin) and cosine (Cos) values ​​are calculated respectively, thereby ensuring the continuity of the time feature in mathematical space.

[0073] As a specific implementation, sine and cosine transforms were performed on the number of hours in a day and the number of days in a week to ensure the continuity and periodicity of the time features in the vector space:

[0074] ; ;

[0075] in, This is the value obtained by sinusoidal transformation of the number of hours in a day. This is the value obtained by cosine transformation of the number of hours in a day. This is the value obtained by sinusoidal transformation of the number of days in a week. This is the value obtained by the cosine transformation of the number of days in a week. This is the original data before the transformation.

[0076] In this embodiment, regarding the construction of time index features, considering the significant periodicity of human activities, not only were discrete numerical features such as hours, days of the week, and months extracted, but periodic encoding technology was also introduced to address the issue of time variables being numerically discontinuous but physically adjacent. Simultaneously, considering the differences in charging behavior between weekdays and weekends, and for morning and evening peak hours, corresponding binary features were constructed to explicitly mark the daily peak and off-peak electricity consumption periods, helping the model capture differences in user travel patterns. Through these methods, structured temporal patterns are extracted from unstructured timestamps, fully leveraging the temporal characteristics within the data.

[0077] Optionally, in some embodiments, the meteorological environmental information includes temperature, humidity, rainfall, and horizontal surface radiation, and the meteorological environmental characteristics include at least one of the following:

[0078] Basic meteorological characteristics, including temperature characteristics corresponding to the temperature, humidity characteristics corresponding to the humidity, rainfall characteristics corresponding to the precipitation, and radiation characteristics corresponding to the horizontal radiation of the land surface;

[0079] The first Boolean feature is used to characterize whether it is hot weather;

[0080] The second Boolean feature is used to characterize whether or not it has rained;

[0081] The second-order interactive features are obtained by performing pairwise dot products on the normalized basic meteorological features.

[0082] In this embodiment, the meteorological environmental features not only include basic meteorological features, but also integrate event-driven discrete signals and high-order interactive information coupled with multiple factors, which significantly improves the model's predictive robustness and accuracy in extreme or complex weather scenarios.

[0083] In some embodiments, the meteorological environmental information includes four raw meteorological observation values: temperature, humidity, precipitation, and surface horizontal radiation. Meteorological environmental characteristics include basic meteorological characteristics, which are constructed based on the meteorological environmental information. Specifically, the basic meteorological characteristics include temperature characteristics corresponding to temperature, humidity characteristics corresponding to humidity, precipitation characteristics corresponding to precipitation, and radiation characteristics corresponding to surface horizontal radiation.

[0084] Optionally, in some embodiments, obtaining meteorological environmental characteristics based on the historical load data includes: constructing a first Boolean feature and a second Boolean feature based on the temperature, humidity, rainfall, and surface horizontal radiation. The first Boolean feature is used to characterize whether it is high-temperature weather, and the second Boolean feature is used to characterize whether there is rainfall. In this embodiment, in order to capture the nonlinear impact of extreme weather on the charging load, Boolean features are constructed. For example, the first Boolean feature used to characterize whether it is high-temperature weather (… ) and the second Boolean feature used to characterize whether rainfall occurs ( These Boolean features, such as , can help the model quickly identify special weather scenarios that may lead to abnormal load fluctuations, making up for the insufficiency of using only continuous variables to characterize abrupt change points.

[0085] In some embodiments, considering that single meteorological factors often exhibit coupling effects (e.g., the pull on cooling load during "hot and humid" sauna-like days is much greater than that of high temperatures alone), second-order interaction features are further constructed. Optionally, in some embodiments, the basic meteorological features are normalized to eliminate dimensional differences, resulting in normalized basic meteorological features. Element-by-element multiplication of these normalized basic meteorological features yields second-order interaction features. For example, the dot product of temperature and humidity features is calculated to obtain the first second-order interaction feature to characterize perceived heat, and the dot product of temperature and radiation features is calculated to obtain the second second-order interaction feature to characterize the overall thermal gain, thereby enhancing the model's sensitivity to extreme weather combinations.

[0086] Given the strong autocorrelation of time series data, historical load is often the best predictor of future load. Optionally, in some embodiments, the basic feature also includes short-term lag features, which include load values ​​from the 1st to the Nth hour prior to the current time, used to characterize the short-term autocorrelation of the load, where N is a positive integer.

[0087] Optionally, in some embodiments, the basic feature further includes a periodic lag feature, which includes the load values ​​corresponding to the previous M periods that have the same periodic phase as the current time, used to capture the cyclic dependence of the load on a daily or weekly scale, where M is a positive integer.

[0088] In this embodiment, multi-scale lag features are constructed, including short-term lag features (such as load in the previous 1 hour and the previous 24 hours) to capture short-term load dependencies, and periodic lag features (such as load at the same time in the previous 7 days and the same time in the previous 14 days) to capture weekly cyclical patterns.

[0089] Specifically, in some embodiments, the specific steps for obtaining the hysteresis characteristics are as follows:

[0090] Using the historical real charging load sequence as the sole benchmark variable, the known historical observations are directly used during the model training phase; while during the online prediction phase, a rolling prediction strategy is adopted, using the actual observations of the previous moment as the input source for subsequent lag features.

[0091] Based on the typical time-dependent structure of power load, a multi-scale lag step set is preset, and different lag step sizes are used to capture load information in different periods.

[0092] For each lag step Perform a time series sliding offset operation: use the Pandas library's shift(k) function to shift the entire load column downwards. a time unit, making the current moment The sample rows are obtained from time 1. The historical load values ​​are used as new feature columns to obtain the lag features.

[0093] It should be understood that strictly ensuring all lag characteristics depend only on Including earlier historical information, it eliminates any future data leaks (look-ahead bias) and ensures the reliability of model evaluation and deployment.

[0094] For example, recent step size (such as ) is used for short-term lags; daily cycle step (e.g.) ) is used for intraday patterns; weekly cycle step size (e.g.) ) is used for weekly cycle modeling.

[0095] Optionally, in some embodiments, the basic features further include rolling statistical features calculated based on multiple preset time windows. The rolling statistical features include at least one of the rolling mean, rolling maximum value, and rolling standard deviation of the load sequence within the preset time window, which are used to characterize the central tendency and dispersion of the load within a local time range.

[0096] In this embodiment, the rolling mean, rolling maximum, and standard deviation for different time windows (such as 6 hours, 12 hours, and 24 hours) were calculated to reflect the recent fluctuation trend and stability of the load sequence.

[0097] For example, in some embodiments, a preset time window is 6 hours. The rolling average of the 6 consecutive hours preceding the current moment is calculated to reflect the average load level over the past half-day, which is sensitive to sudden events (such as the end of a large event or the concentrated activation of charging stations). The rolling maximum value of the 6 consecutive hours preceding the current moment is calculated to capture short-term peak loads and identify the risk of abnormally high loads. The rolling standard deviation of the 6 consecutive hours preceding the current moment is calculated to measure the severity of load fluctuations within 6 hours. The larger the standard deviation, the more irregular the user's charging behavior or the stronger the influence of external disturbances.

[0098] In some embodiments, the preset time window is 12 hours to cover a half-day cycle (e.g., 6 a.m. to 6 p.m.). In this embodiment, the rolling average can characterize the typical load baseline during the day or night, the rolling maximum value can identify the highest demand within half a day to assist capacity planning, and the rolling standard deviation can assess the stability of half-day operation, such as whether there are sustained high fluctuations in the afternoon of a weekday.

[0099] In some embodiments, the preset time window is 24 hours to cover a complete calendar day. In this embodiment, the calculated rolling average can represent the "overall load level of yesterday" as an important reference benchmark for today's forecast; while the calculated rolling maximum can reflect the peak load of the previous day and is used to determine whether it is close to the system capacity limit; the calculated rolling standard deviation is used to quantify the dispersion of the load distribution throughout the day, with a low standard deviation indicating a flat load curve (such as on holidays), and a high standard deviation corresponding to a typical "double-peak" workday pattern.

[0100] Optionally, in some embodiments, obtaining prior features based on the historical load data includes:

[0101] The load value is decomposed into a combination of trend, seasonal, holiday, and error terms using the Prophet model.

[0102] The value of the trend term at each time point is extracted as the trend component, the seasonal component is determined based on the seasonality term, and the holiday component is determined based on the holiday term.

[0103] Predicted value components are generated based on the trend item, the seasonality item, and the holiday item;

[0104] Determine the prior features, which include the trend component, the seasonal component, the holiday component, and the predicted value component.

[0105] In this embodiment, the Prophet model is constructed as a pure time-series prior feature extractor, aiming to decouple rigid time patterns from historical data and provide a robust benchmark for subsequent ensemble models. The Prophet model first decomposes the time series into a long-term trend term and multiple seasonal terms. The trend term is fitted using a piecewise linear model, which can dynamically capture the load baseline drift caused by the growth of electric vehicle ownership. The seasonal terms use Fourier series to characterize the periodic fluctuations at different time scales such as daily, weekly, and yearly.

[0106] It should be understood that the input to the Prophet model is strictly limited to time indices and historical loads. To avoid functional overlap with subsequent prediction models, the input to the Prophet model does not include external covariates such as meteorological and environmental characteristics. Specifically, the input vector includes:

[0107] A timestamp sequence, a standardized time series index, in the format YYYY-MM-DD HH:MM:SS.

[0108] Historical load values, which are the actual power load values ​​at the corresponding points in time, only include historical data up to the prediction point in the rolling verification.

[0109] Holiday information, including a list of Chinese statutory holidays and their corresponding workday adjustments.

[0110] First, the Facebook Prophet model is introduced to decompose the time-series characteristics of charging load. Based on the specific fluctuation characteristics of the charging load data, the Prophet model in this embodiment can be flexibly configured as an additive model or a multiplicative model to adapt to the load evolution patterns under different scenarios. This embodiment achieves adaptive modeling of the entire lifecycle of charging load (from the initial stable period to the rapid expansion period) by flexibly switching between the additive and multiplicative architectures of the Prophet model, significantly improving the robustness and accuracy of charging station load prediction at different development stages.

[0111] In some embodiments, when the seasonal fluctuations in charging load are relatively constant and do not significantly amplify with the increase of the overall baseline trend, an additive model is used. (The text then abruptly shifts to a different topic: "Time...") load value It can be decomposed into a linear sum of trend, seasonality, holiday, and error terms, and its mathematical expression is:

[0112] ;

[0113] In this embodiment, an additive model can be used for decomposition when the charging infrastructure is in its early stages of development and the overall load base is small and fluctuates smoothly. In scenarios where the total load is small and the fluctuation amplitude is constant, the additive model avoids the problem of excessive amplification or compression of periodic components caused by small changes in the base, ensuring that the estimation of seasonal factors and holiday factors is more robust and effectively reducing the risk of overfitting caused by data sparsity.

[0114] In other embodiments, when the charging load exhibits significant heteroscedasticity, meaning that the periodic fluctuations in load (such as the difference between morning and evening peak hours) increase proportionally with the overall number of electric vehicles, the specific configuration is a multiplicative seasonality mode (seasonality_mode='multiplicative'). In this case, the model will time... The load value is decomposed into a combination of the trend term and the product of other periodic terms, and its mathematical expression is:

[0115] ;

[0116] In this embodiment, a multiplicative model is used to process charging networks in a rapid growth phase with significant heteroscedasticity, dynamically adapting to the evolution of fluctuation amplitude as the trend increases. When facing future scenarios where the load base changes drastically, the multiplicative model can more accurately extrapolate future load peaks based on the learned relative seasonality without retraining.

[0117] in, The trend term is used to fit the non-periodic baseline drift caused by the growth of electric vehicle ownership and changes in user habits. In some embodiments, this drift typically exhibits a slow upward long-term trend, but may show a slowdown in growth, a plateau, or even a period of saturation at certain stages (e.g., affected by the reduction of policy subsidies, grid capacity limitations, or market penetration approaching its upper limit). Therefore, a trend function with saturation modeling capabilities or abrupt change adaptation capabilities is required.

[0118] In some embodiments, to characterize the asymptotic saturation characteristics of load growth, the trend term The Logistic Growth Model is adopted, which naturally fits an S-shaped curve of rapid initial growth, slowing growth in the middle stage, and stabilization in the later stage. This conforms to the evolution of electric vehicle charging load in a limited market space and avoids the risk of overestimation caused by linear extrapolation. Its specific mathematical expression is as follows:

[0119] ;

[0120] in, Capacity is a time-varying capacity that represents the upper limit of the grid capacity or the number of electric vehicles in the region. The base growth rate determines the steepness of the S-curve; The offset parameter represents the time point when the curve reaches its maximum growth rate (i.e., half of the upper limit of saturation).

[0121] In some embodiments, when there are structural abrupt changes in the historical load sequence (such as large-scale operation of new urban areas, implementation of major policies, or changes in travel patterns after the pandemic), the trend term adopts a piecewise linear trend model. The model automatically detects potential change points in the time series. Typically, it is set to uniformly distribute several candidate points on the time axis (e.g., one every 25 days). By introducing L1 regularization priors, sparse constraints are applied to the slope (growth rate) of each segment, activating change points only when the data significantly supports them. The final generated... For a piecewise continuous and piecewise linear function, abrupt changes in slope are allowed at turning points, but the function values ​​remain continuous. The trend component is obtained by extracting the value of the trend term at each time step. ].

[0122] For the seasonality term, specifically, a Fourier series is used to fit the multi-scale periodic fluctuations of the load on weekly and yearly scales. Optionally, in some embodiments, the seasonal component includes daily periodic characteristic components and weekly periodic characteristic components, and the extraction of the values ​​of the seasonality term at each time point as the seasonal component includes:

[0123] The seasonality term is decomposed into multi-scale components to extract the daily and weekly characteristic components, respectively. The daily characteristic components are fitted with load waveforms of morning and evening peaks and nighttime troughs based on a 24-hour Fourier series. The weekly characteristic components are fitted with load differences between weekdays and weekends based on a 7-day Fourier series.

[0124] For the daily cycle, capture the 24-hour waveform of morning and evening peaks and nighttime troughs (set daily_seasonalirty=True) to obtain the daily cycle characteristic components. For the weekly cycle, the load difference between weekdays and weekends is captured to obtain the weekly cycle characteristic components. It should be understood that, given the data span limitations, yearly fitting is not enabled in this embodiment to avoid overfitting.

[0125] In some embodiments, the seasonal item can be represented as:

[0126] .

[0127] in, The order of the Fourier series is used to control the smoothness and complexity of the fitted periodic curve. For a preset seasonal cycle length (e.g., when fitting a daily cycle) Hours, when fitting the period sky); and The Fourier coefficients are obtained by optimizing the fit of the model using historical data. This is the corresponding time index variable.

[0128] This is a holiday term used to capture load mutations during statutory holidays and special events. Specifically, holidays are treated as independent shock events, and the model learns an independent impact coefficient for each holiday. via indicator function By superimposing this onto the model, the increase or decrease in workload on a specific date compared to a regular workday is quantified, resulting in a holiday component. ].

[0129] In some embodiments, holiday items can be represented as:

[0130] ;

[0131] in, The total number of categories of holidays is preset; For the first The set of dates corresponding to various holidays; For the indicator function, at the current time The value is 1 if it belongs to the set of holidays, and 0 otherwise. This represents the corresponding influence coefficient.

[0132] This is a random error term that follows a normal distribution. Using this model, this embodiment extracts a comprehensive fitted value containing trend and periodic information. For example, in some embodiments, based on the heteroscedasticity of the load data, a multiplicative combination method is used to generate the comprehensive fitted value as a component of the predicted value. ]:

[0133] ;

[0134] in, This is a daily periodic characteristic component. As a periodic feature component, this predicted value component is a "meta-feature" that integrates trends, all seasonal cycles, and holiday effects. It highly condenses complex time-series patterns into a single value, essentially providing a high-precision initial benchmark for subsequent prediction models. This allows the model to focus on fitting the remaining residual fluctuations without having to learn the basic patterns from scratch.

[0135] In some embodiments, the Prophet model outputs the following core feature components as prior features. Trend components [ Date feature component ], periodic characteristic components[ ], Holiday portion ], predicted value components [ ].

[0136] Table 1 Typical Characteristic Components Table

[0137]

[0138] In some embodiments, the basic features include time index features, meteorological environment features, short-term lag features, periodic lag features, and rolling statistical features. In this embodiment, the feature pool has been expanded to include a feature space of over one hundred dimensions, including meteorology, time, Prophet decomposition, lag terms, and statistics (some typical features are shown in Table 1 above).

[0139] However, high-dimensional feature spaces not only increase the computational cost of the model but may also introduce noise and multicollinearity, leading to overfitting. As a specific implementation, a two-stage feature selection strategy is used to screen features based on both fundamental and prior features. The first stage is initial correlation screening, calculating the Pearson correlation coefficient between all candidate features and the target variable (actual load), eliminating weak features with correlations below a threshold (e.g., 0.2), quickly filtering out invalid information. The second stage is redundancy control, using a greedy selection based on the cross-correlation matrix between features. A redundancy threshold (e.g., 0.9) is set; when a new feature's correlation with any feature in the selected feature set exceeds this threshold, the feature is discarded. This strategy effectively avoids the simultaneous selection of highly similar features such as "1-hour lag" and "2-hour lag," ensuring that the feature set maintains high information content while possessing good independence.

[0140] As a specific implementation, after dual screening based on feature redundancy and correlation, 15 optimal features were finally identified, as detailed in Table 2. Correlation analysis shows that lag features dominate in explaining load fluctuations, with the actual load value lagging by one day (…) With a correlation coefficient of 0.9740, this data tops the list, revealing the strong diurnal periodicity of electric vehicle (EV) charging load. Following closely is the load data at the same time, lagging 7 days, reflecting weekly historical patterns. and load with a 2-hour lag characterizing short-term dependence .

[0141] Table 2 Optimal Feature Set

[0142]

[0143] It is worth noting that the extraction based on Prophet decomposition... The features also demonstrated good predictive ability. This effectively supplements the nonlinear trend information that traditional statistical features struggle to capture. Furthermore, hourly temperature ( ) and whether it is high temperature ( The inclusion of environmental interaction features, such as [missing information], further enhances the model's adaptability to extreme weather scenarios. The final feature subset maintains high relevance while controlling the maximum feature redundancy at 0.8960 (<0.9), ensuring both sparsity and efficiency of the feature matrix.

[0144] As a specific implementation example, as shown in Table 2, by transforming these decomposed components into explicit numerical features and combining them with lag components and rolling statistics, a high-dimensional feature set containing 15 key components was constructed. Correlation analysis showed that the correlation coefficient between the predicted value components extracted by Prophet and the actual load was as high as 0.97, significantly enriching the model's input information and greatly reducing the difficulty of subsequent supervised learning tasks.

[0145] In some embodiments, fusing the prior features and the basic features to obtain the input feature vector includes:

[0146] The prior features and the basic features are initially screened for correlation. The Pearson correlation coefficient between all candidate features and the actual load is calculated. Weak features with correlation below the threshold are removed. The selected features are then placed into the selected feature set.

[0147] A greedy selection is performed based on the cross-correlation matrix between features. A redundancy threshold is set. When the correlation between a new feature and any feature in the selected feature set exceeds the threshold, the feature will be discarded, and the optimal feature will be obtained in the end.

[0148] The optimal features obtained from the screening are fused to obtain the input feature vector.

[0149] It should be understood that the specific structure of the prediction module is not limited here. For example, in some embodiments, the prediction module is built based on a random forest. In other embodiments, the prediction module is built based on an LSTM network. In still other embodiments, the prediction module is built based on a Transformer.

[0150] Experimental results confirm that although deep learning models have theoretical advantages in temporal capture, XGBoost achieves the highest prediction accuracy (R²=0.9915) and the lowest computational cost, provided that feature engineering is sufficient, thanks to its strong fitting ability to tabular data and robustness to outliers.

[0151] Optionally, in some embodiments, the prediction module is built based on the XGBoost model. In this embodiment, XGBoost is selected as the final regression predictor. XGBoost is an efficient ensemble algorithm based on gradient boosting decision trees, which optimizes the objective function through serial iteration to approximate the true load.

[0152] In the In the round of iteration, its objective function It consists of a loss function and a regularization term:

[0153]

[0154] in, A differentiable convex loss function to measure prediction error; Indicates the first Tree samples The predicted increment, For the first The actual load observation value of each sample For the front After round of iterations, the model is on the first The cumulative predicted value of each sample. For the current number Round-generated decision tree model, The total number of samples in the input training set; This is a regularization term used to penalize the complexity of the tree (such as the number of leaf nodes and leaf weights) to prevent overfitting. Compared to traditional statistical models, XGBoost has extremely strong feature interaction mining capabilities, effectively handling the nonlinear coupling relationship between Prophet prior features, time index features, and meteorological environment features.

[0155] In some embodiments, the final input feature vector is constructed. This vector incorporates time index features. Meteorological and environmental characteristics Historical lag characteristics and the prior features extracted by Prophet ,Right now:

[0156] ;

[0157] Final prediction results Output of the trained XGBoost model:

[0158] .

[0159] In this embodiment, the Prophet model does not directly output the final prediction result, but rather acts as a high-order feature extractor. Furthermore, the Prophet model does not input any meteorological or environmental information or other cofactors, only extracting baseline features using pure time series data. The prior features obtained by the Prophet model are used as prior knowledge and input into the prediction model. The prediction model specifically learns the residuals that the Prophet model could not explain (caused by nonlinear meteorological factors or sudden events), thus achieving higher accuracy than simple weighted fusion. The sensitivity of power load to temperature is typically nonlinear (e.g., a U-shaped curve). Prophet's built-in linear regression cannot capture this complex relationship. In this embodiment, meteorological and environmental features are processed by the prediction model, which can effectively learn the nonlinear threshold effects and interaction effects of environmental factors.

[0160] The following describes the specific implementation steps of the charging load prediction method based on time-series decomposition provided by the present invention, using a specific embodiment as an example.

[0161] First, in the data input and preprocessing stage, the original charging load records and meteorological observation data are preprocessed, undergoing rigorous data cleaning and integration to obtain historical load data. Specifically, preprocessing includes imputing missing values ​​using the mean of adjacent time periods, correcting outlier zero values ​​based on historical distribution, and performing time alignment and fusion of multi-source heterogeneous data at the hourly granularity to ensure data continuity and consistency. Based on this, a basic feature set (OriginalFeatures, OF) is constructed, including time index features, meteorological environment features, short-term lag features, periodic lag features, and rolling statistical features, providing basic physical and statistical information input for the model.

[0162] Secondly, a time-series decomposition and feature enhancement based on Prophet is implemented. The additive architecture of the Prophet model is used to deconstruct the non-stationary load series into long-term trends, multiple seasonalities, and holiday effects. The comprehensive predictions generated by Prophet are extracted as prediction components and deeply fused with the basic feature set. To address the noise and multicollinearity issues arising from the high-dimensional feature space, a two-stage feature selection strategy is designed: the first stage uses the Pearson correlation coefficient to eliminate weakly correlated features; the second stage sets a threshold based on the cross-correlation matrix for redundancy control, ultimately selecting the optimal feature subset that combines high information content with low redundancy.

[0163] Finally, XGBoost was selected as the final regression predictor. The input feature vector was obtained by feature fusion based on the optimal feature subset. The input feature vector was then input into the prediction module built on XGBoost for load prediction, and the final prediction result was obtained.

[0164] In this embodiment, the advantages of Prophet in handling structural patterns and the ability of XGBoost to handle nonlinear residuals are fully utilized, ensuring the prediction accuracy and generalization performance of the model from a mechanistic perspective, and effectively solving the problem that traditional methods are unable to capture nonlinear periodic patterns. Unlike existing studies that rely solely on raw meteorological data and simple time indexes, this application innovatively uses the Prophet model as a feature extractor, utilizing its additive model architecture to deconstruct complex load sequences into long-term trends, multiple seasonal cycles (weekly, annual), and holiday effects.

[0165] In some embodiments, in long-term time series prediction tasks, traditional static dataset partitioning methods (e.g., using the first 80% as the training set and the last 20% as the test set) often assume that the data distribution remains stable over time. However, electric vehicle charging load is significantly affected by seasonal climate change and socio-economic activities, exhibiting a clear "seasonal concept drift" phenomenon. If static partitioning is used directly, the model may only be trained based on the high-temperature, high-load characteristics of summer, while its generalization ability under the low-temperature scenario of winter will decrease significantly. Furthermore, the rapid growth in the number of electric vehicles leads to a continuous rise in the load baseline, making it difficult for static partitioning to enable the model to adapt to the latest load growth trends in a timely manner.

[0166] To address the aforementioned issues, some implementations replace the traditional static hold-out method with a monthly rolling validation strategy. Monthly rolling validation divides the entire year's time-series data into 12 independent subsets by month, and performs rolling training and evaluation within each monthly window. The specific partitioning logic is as follows:

[0167] Let the full dataset for month m be... Divide it into training sets according to time sequence. Validation set With test set Specifically, Data from the first 20 to 23 days of the month, covering the main weekdays and weekends, was selected to allow the model to fully learn the typical daily load curve shape and meteorological response patterns. The following four days of data constituted the validation set. This is primarily used to guide the fine-tuning of prediction model hyperparameters (such as tree depth and learning rate) and to trigger early stopping mechanisms to prevent overfitting. The remaining days of the month are used as the test set. This portion of data is strictly isolated and used solely to evaluate the model's final predictive performance in the current month's environment. The mathematical expression of this strategy can be summarized as follows: for any prediction time, the corresponding model is obtained solely through training.

[0168] To visually verify the effectiveness of the aforementioned monthly rolling partitioning strategy in addressing seasonal concept drift, this example compares the evolution trend of this strategy with that of traditional static partitioning (80% training, 20% testing) in terms of prediction error throughout the year. For details, please refer to [link to relevant documentation]. Figure 2 Rolling partitioning strategy ( Figure 2 The red curve (in the middle) shows a significant error convergence characteristic. In the early stages of forecasting (e.g., February to April), due to insufficient historical sample accumulation (i.e., the "cold start" phase), the model struggles to fully capture complex load fluctuation patterns, resulting in a relatively high Mean Absolute Percentage Error (MAPE). However, as time progresses and the training window is continuously updated, the model incorporates the latest meteorological and load characteristics, leading to a steady improvement in forecast accuracy. Notably, in the fourth quarter, when data accumulation is sufficient, such as in November and December, the MAPE of the rolling partition model is significantly lower than the baseline of the static partition. This result strongly demonstrates that the rolling partition strategy effectively overcomes seasonal distribution bias in long-term forecasts by dynamically adapting to changes in data distribution, exhibiting stronger robustness and generalization ability compared to static partitioning.

[0169] In this embodiment, a monthly rolling partitioning strategy is proposed to address the seasonal distribution bias in power load forecasting. This strategy ensures that the training data for each forecast fully covers the latest meteorological characteristics and load growth trends of the current season, thereby effectively overcoming the impact of data distribution drift and making the evaluation results more objectively reflect the robustness of the model across different seasons throughout the year.

[0170] like Figure 3As shown, this embodiment of the invention also provides a charging load prediction device 300 based on time-series decomposition, comprising:

[0171] The first acquisition module 301 is used to acquire historical load data, which includes load values ​​and meteorological information within a preset period.

[0172] The second acquisition module 302 is used to acquire prior features and basic features based on the historical load data. The basic features include time index features and meteorological environment features. The prior features are obtained by decomposing the time series characteristics of the historical load data using the Prophet model. The time index features are determined based on time information within the preset period. The meteorological environment features are determined based on the meteorological environment information.

[0173] The fusion module 303 is used to fuse the prior features and the basic features to obtain the input feature vector;

[0174] The input module 304 is used to input the input feature vector into a pre-trained prediction module for load prediction and obtain the prediction result.

[0175] Optionally, obtaining prior features based on the historical load data includes:

[0176] The load value is decomposed into a combination of trend, seasonality, holiday, and error terms using the Prophet model. The trend term is used to fit the non-periodic baseline drift, the seasonality term is used to fit the periodic fluctuations of the load value on the weekly and yearly scales, the holiday term is used to capture load abrupt changes during statutory holidays, and the error term is a random error that follows a normal distribution.

[0177] The value of the trend item at each time point is extracted as a trend component, the value of the seasonality item at each time point is extracted as a seasonal component, the value of the holiday item at each time point is extracted as a holiday component, and a predicted value component is generated based on the trend item, the seasonality item, and the holiday item.

[0178] Determine prior features, which include at least one of the trend component, the seasonal component, the holiday component, and the predicted value component.

[0179] Optionally, the seasonal component includes a daily periodic feature component and a weekly periodic feature component, and the extraction of the seasonal term's value at each time point as the seasonal component includes:

[0180] The seasonality term is decomposed into multi-scale components to extract the daily periodic feature components and the weekly periodic feature components, respectively.

[0181] The daily periodic characteristic component is based on the load waveform of morning and evening peaks and nighttime troughs fitted by a 24-hour Fourier series; the weekly periodic characteristic component is based on the load difference between weekdays and weekends fitted by a 7-day Fourier series.

[0182] Optionally, the time step can be determined using the Prophet model. load value The breakdown into trend, seasonality, holiday, and error components is as follows:

[0183] ;

[0184] Alternatively, the time step can be determined using the Prophet model. load value The breakdown into trend, seasonality, holiday, and error components is as follows:

[0185] ;

[0186] in, For the trend term, For the aforementioned seasonal item, For the aforementioned holiday items, This refers to the error term.

[0187] Optionally, the prediction module is built based on the XGBoost model.

[0188] Optionally, the basic features also include at least one of the following:

[0189] Short-term lag characteristics, which include load values ​​from the 1st to the Nth hour prior to the current time, are used to characterize the short-term autocorrelation of the load, where N is a positive integer;

[0190] The periodic lag feature includes the load values ​​at the M previous periods that have the same periodic phase as the current time, which are used to capture the cyclic dependence of the load on a daily or weekly scale, where M is a positive integer.

[0191] The rolling statistical features are calculated based on multiple preset time windows. The rolling statistical features include at least one of the rolling mean, rolling maximum value and rolling standard deviation of the load sequence within the preset time window, which are used to characterize the central tendency and dispersion of the load within a local time range.

[0192] Optionally, the meteorological environmental information includes temperature, humidity, rainfall, and surface horizontal radiation, and the meteorological environmental characteristics include at least one of the following:

[0193] Basic meteorological characteristics, including temperature characteristics corresponding to the temperature, humidity characteristics corresponding to the humidity, rainfall characteristics corresponding to the precipitation, and radiation characteristics corresponding to the horizontal radiation of the land surface;

[0194] The first Boolean feature is used to characterize whether it is hot weather;

[0195] The second Boolean feature is used to characterize whether or not it has rained;

[0196] The second-order interactive features are obtained by performing pairwise dot products on the normalized basic meteorological features.

[0197] The charging load prediction device 300 based on time-series decomposition provided in this application embodiment can execute the above method embodiment, and its implementation principle and technical effect are similar, so it will not be described again here.

[0198] It should be noted that the division of units in the embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0199] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0200] like Figure 4As shown, this application embodiment provides an electronic device 400, including: a memory 402, a processor 401, and a program stored in the memory 402 and executable on the processor 401; the processor 401 is used to read the program in the memory 402 to implement the steps in the charging load prediction method based on time-series decomposition as described above.

[0201] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the various processes of the above-described charging load prediction method based on time-series decomposition and achieves the same technical effect. To avoid repetition, it will not be described again here. The readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (such as compact disks (CDs), digital video discs (DVDs), Blu-ray discs (BDs), high-definition versatile discs (HVDs), etc.), and semiconductor storage (such as read-only memory (ROMs), erasable programmable read-only memory (EPROMs), electrically erasable programmable read-only memory (EEPROMs), non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0202] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0203] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0204] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of this application without departing from its spirit, and all of these modifications are within the scope of protection of this application.

Claims

1. A charging load prediction method based on time-series decomposition, characterized in that, include: Acquire historical load data, which includes load values ​​and meteorological information within a preset period; Based on the historical load data, prior features and basic features are obtained. The basic features include time index features and meteorological environment features. The prior features are obtained by decomposing the time series characteristics of the historical load data using the Prophet model. The time index features are determined based on time information within the preset period. The meteorological environment features are determined based on the meteorological environment information. The prior features and the basic features are fused to obtain the input feature vector; The input feature vector is input into a pre-trained prediction module to perform load prediction and obtain the prediction result. The step of obtaining prior features based on the historical load data includes: The load value is decomposed into a combination of trend, seasonality, holiday, and error terms using the Prophet model. The trend term is used to fit the non-periodic baseline drift, the seasonality term is used to fit the periodic fluctuations of the load value on the weekly and yearly scales, the holiday term is used to capture load abrupt changes during statutory holidays, and the error term is a random error that follows a normal distribution. The value of the trend item at each time point is extracted as a trend component, the value of the seasonality item at each time point is extracted as a seasonal component, the value of the holiday item at each time point is extracted as a holiday component, and a predicted value component is generated based on the trend item, the seasonality item, and the holiday item. Determine prior features, which include at least one of the trend component, the seasonal component, the holiday component, and the predicted value component; The prediction module is built based on the extreme gradient boosting XGBoost model. The meteorological environmental information includes temperature, humidity, rainfall, and horizontal surface radiation, and the meteorological environmental characteristics include at least one of the following: Basic meteorological characteristics, including temperature characteristics corresponding to the temperature, humidity characteristics corresponding to the humidity, rainfall characteristics corresponding to the precipitation, and radiation characteristics corresponding to the horizontal radiation of the land surface; The first Boolean feature is used to characterize whether it is hot weather; The second Boolean feature is used to characterize whether or not it has rained; The second-order interactive features are obtained by performing pairwise dot products on the normalized basic meteorological features.

2. The charging load prediction method based on time-series decomposition according to claim 1, characterized in that, The seasonal component includes a daily periodic feature component and a weekly periodic feature component. Extracting the values ​​of the seasonal term at each time point as the seasonal component includes: The seasonality term is decomposed into multi-scale components to extract the daily periodic feature components and the weekly periodic feature components, respectively. The daily periodic characteristic component is based on the load waveform of morning and evening peaks and nighttime troughs fitted by a 24-hour Fourier series; the weekly periodic characteristic component is based on the load difference between weekdays and weekends fitted by a 7-day Fourier series.

3. The charging load prediction method based on time-series decomposition according to claim 1, characterized in that, Using the Prophet model to determine the time... load value The breakdown into trend, seasonality, holiday, and error components is as follows: ； Alternatively, the time step can be determined using the Prophet model. load value The breakdown into trend, seasonality, holiday, and error components is as follows: ； in, For the trend term, For the aforementioned seasonal item, For the aforementioned holiday items, This refers to the error term.

4. The charging load prediction method based on time-series decomposition according to claim 1, characterized in that, The fundamental features also include at least one of the following: Short-term lag characteristics, which include load values ​​from the 1st to the Nth hour prior to the current time, are used to characterize the short-term autocorrelation of the load, where N is a positive integer; The periodic lag feature includes the load values ​​at the M previous periods that have the same periodic phase as the current time, which are used to capture the cyclic dependence of the load on a daily or weekly scale, where M is a positive integer. The rolling statistical features are calculated based on multiple preset time windows. The rolling statistical features include at least one of the rolling mean, rolling maximum value and rolling standard deviation of the load sequence within the preset time window, which are used to characterize the central tendency and dispersion of the load within a local time range.

5. A charging load prediction device based on time-series decomposition, characterized in that, include: The first acquisition module is used to acquire historical load data, which includes load values ​​and meteorological information within a preset period. The second acquisition module is used to acquire prior features and basic features based on the historical load data. The basic features include time index features and meteorological environment features. The prior features are obtained by decomposing the time series characteristics of the historical load data using the Prophet model. The time index features are determined based on time information within the preset period. The meteorological environment features are determined based on the meteorological environment information. The fusion module is used to fuse the prior features and the basic features to obtain the input feature vector; The input module is used to input the input feature vector into the pre-trained prediction module for load prediction and obtain the prediction result; The step of obtaining prior features based on the historical load data includes: The load value is decomposed into a combination of trend, seasonality, holiday, and error terms using the Prophet model. The trend term is used to fit the non-periodic baseline drift, the seasonality term is used to fit the periodic fluctuations of the load value on the weekly and yearly scales, the holiday term is used to capture load abrupt changes during statutory holidays, and the error term is a random error that follows a normal distribution. The value of the trend item at each time point is extracted as a trend component, the value of the seasonality item at each time point is extracted as a seasonal component, the value of the holiday item at each time point is extracted as a holiday component, and a predicted value component is generated based on the trend item, the seasonality item, and the holiday item. Determine prior features, which include at least one of the trend component, the seasonal component, the holiday component, and the predicted value component; The prediction module is built based on the extreme gradient boosting XGBoost model. The meteorological environmental information includes temperature, humidity, rainfall, and horizontal surface radiation, and the meteorological environmental characteristics include at least one of the following: Basic meteorological characteristics, including temperature characteristics corresponding to the temperature, humidity characteristics corresponding to the humidity, rainfall characteristics corresponding to the precipitation, and radiation characteristics corresponding to the horizontal radiation of the land surface; The first Boolean feature is used to characterize whether it is hot weather; The second Boolean feature is used to characterize whether or not it has rained; The second-order interactive features are obtained by performing pairwise dot products on the normalized basic meteorological features.

6. An electronic device, comprising: A memory, a processor, and a program stored in the memory and executable on the processor; characterized in that the processor is configured to read the program from the memory to implement the steps in the charging load prediction method based on time-series decomposition as described in any one of claims 1 to 4.

7. A readable storage medium for storing a program, characterized in that, When the program is executed by the processor, it implements the steps in the charging load prediction method based on time-series decomposition as described in any one of claims 1 to 4.

Images