Rural power distribution network load forecasting method and device considering electric vehicle user behavior

Abstract

This application discloses a method and apparatus for rural power distribution network load forecasting that considers electric vehicle user behavior, belonging to the field of power system load forecasting technology. The method includes: collecting four types of multi-source raw data from the target area: historical load, electricity price policy, agricultural activities, and meteorology; constructing a continuous numerical electricity price feature sequence, a dynamic agricultural busyness coefficient sequence integrating agricultural activities and electric vehicle charging records, and a meteorological feature sequence; aligning the various feature time series with the load time series, filtering key features based on correlation, and concatenating them into a multi-dimensional user behavior feature vector; feeding the vector into a deep learning model for training, and predicting future load and outputting the results after the model converges. By integrating electricity price, agricultural activities, and meteorology to characterize rural electric vehicle charging behavior, refined features effectively improve load forecasting accuracy; time series alignment and feature filtering reduce redundant information; deep learning adapts to time series changes; and the model exhibits excellent generalization ability, providing accurate data support for rural power distribution network planning and orderly charging.

Description

Technical Field

[0001] This application relates to the field of power system load forecasting technology, and in particular to a method and apparatus for forecasting rural distribution network load considering the behavior of electric vehicle users. Background Technology

[0002] With the deepening of the "dual carbon" goals and the implementation of policies to promote new energy vehicles in rural areas, the number of electric vehicles in counties is showing an increasing trend. Electric vehicle charging loads are characterized by strong spatiotemporal randomness and are greatly affected by user behavior, thus their connection has a certain impact on the safe and stable operation of rural power distribution networks.

[0003] Traditional load forecasting methods typically use historical load data as the primary input, failing to adequately consider the impact of unique user behaviors in county-level areas, such as agricultural activity rhythms and time-of-use pricing responses, on charging load. Existing load forecasting studies involving user behavior primarily focus on urban transportation, lacking in-depth modeling of travel patterns caused by agricultural activities in county-level areas. Regarding the quantification of external characteristics, some studies treat electricity prices as discrete categorical variables, failing to fully utilize the continuous guiding strength information of price signals; and for characterizing agricultural activity factors, most studies use binary representations, which struggle to reflect the gradual changes in agricultural activity levels and intraday fluctuations. These factors, to some extent, affect the forecasting accuracy of load forecasting methods in county-level scenarios. Summary of the Invention

[0004] The main objective of this application is to provide a method and apparatus for predicting rural power distribution network load that takes into account the behavior of electric vehicle users, in order to solve the technical problem that existing rural power distribution network load prediction methods do not fully consider user behavior factors such as electricity pricing policies and agricultural activities, resulting in insufficient prediction accuracy.

[0005] To achieve the above objectives, this application proposes a rural power distribution network load forecasting method that considers the behavior of electric vehicle users. The method includes: Acquire multi-source raw data for the target area, including historical load time series data, electricity price policy data, agricultural activity data, and meteorological observation data; A continuous numerical electricity price characteristic sequence was constructed based on electricity price policy data; a busy farming coefficient sequence representing the intensity of busy farming was constructed based on agricultural activity data and historical charging records of electric vehicles; and a meteorological characteristic sequence was extracted from meteorological observation data. The electricity price feature sequence, the agricultural busyness coefficient sequence, and the meteorological feature sequence are aligned with the historical load time series data in terms of time dimension. Based on correlation analysis, key features related to the load are selected, and the selected key features are spliced ​​with the historical load time series data to form a multi-dimensional user behavior feature vector. The multidimensional user behavior feature vector is input into a pre-built deep learning load prediction model for training. After the model converges, the load prediction model is used to predict the load at future times and output the load prediction results.

[0006] In one embodiment, the step of constructing a continuous numerical sequence of electricity price characteristics includes: The electricity price period for each time point is determined based on the electricity price policy data, and the electricity price value corresponding to each time point is used as the basic electricity price component. Based on the base electricity price component, the price difference between adjacent time points is extracted as the electricity price change trend component; Extract the ratio between the current electricity price and the lowest electricity price of the day as the relative level component of the electricity price; The basic electricity price component, the electricity price change trend component, and the electricity price relative level component are combined to form a continuous numerical electricity price feature sequence.

[0007] In one embodiment, the step of constructing a sequence of agricultural busyness coefficients characterizing the intensity of agricultural busyness includes: The monthly baseline agricultural busyness coefficient for each month is determined based on agricultural activity data; The daily time periods are divided according to the travel and work patterns of rural residents, and a daily fluctuation coefficient is assigned to each daily time period. Multiply the monthly baseline agricultural busyness coefficient by the intraday fluctuation coefficient of the corresponding date to obtain the stratified agricultural busyness coefficient, which serves as the base value for the agricultural busyness coefficient series. The system acquires historical charging records of electric vehicles, extracts charging behavior features from these records, and performs data-driven dynamic calibration of the hierarchical agricultural busyness coefficients based on these features, generating a calibrated agricultural busyness coefficient sequence.

[0008] In one embodiment, the step of data-driven dynamic calibration of the stratified agricultural busyness coefficient based on charging behavior characteristics includes: Extract the charging frequency, average charging amount, concentration of charging time period, and average charging duration from historical charging records as charging behavior characteristics. Input the charging behavior characteristics into a preset mapping function, and output the initial estimate of the busy farming intensity derived from the charging behavior; The initial estimate of the busy farming season intensity is fused with the prior busy farming season intensity determined based on agricultural activity data to obtain the calibrated busy farming season intensity, which is then used to update the stratified busy farming season coefficient.

[0009] In one embodiment, the step of concatenating the selected key features with historical load time-series data to form a multidimensional user behavior feature vector includes: The Spearman correlation coefficient method was used to calculate the correlation measure between the electricity price characteristic sequence, the agricultural busyness coefficient sequence, the meteorological characteristic sequence and the historical load time series data, respectively. Features whose relevance measure is greater than a preset threshold are identified as key features; The key features are aligned with historical load time series data in the time dimension and concatenated by time step so that each time step contains the load value and its corresponding key feature value, forming a multi-dimensional user behavior feature vector.

[0010] In one embodiment, the deep learning load prediction model is a long short-term memory network model. The step of training the pre-built deep learning load prediction model by inputting multi-dimensional user behavior feature vectors includes: Normalize the multidimensional user behavior feature vectors to eliminate the differences in the dimensions of each feature; A training sample set is constructed from the normalized multidimensional user behavior feature vector using a sliding window approach; The training sample set is input into the Long Short-Term Memory (LSTM) network model, which consists of an input layer, an LSTM hidden layer, a Dropout layer, a fully connected layer, and an output layer connected in sequence. An optimizer combined with a gradient pruning strategy is used to optimize the parameters of the long short-term memory network model, and the model training is completed when the convergence condition is met.

[0011] In one embodiment, after the step of outputting the load forecast result, the method further includes: Acquire multi-source raw data under various test scenarios, including at least different seasonal scenarios, different date type scenarios, and different electricity price time periods; The multi-source raw data from each test scenario are input into the trained deep learning load prediction model to obtain the load prediction results for each test scenario. Based on the load prediction results for each test scenario and the actual load data for that scenario, the generalization ability and robustness of the deep learning load prediction model are evaluated.

[0012] Furthermore, to achieve the above objectives, this application also proposes a rural power distribution network load forecasting device that considers the behavior of electric vehicle users. The rural power distribution network load forecasting device considering the behavior of electric vehicle users includes: The data acquisition module is used to acquire multi-source raw data of the target area, including historical load time series data, electricity price policy data, agricultural activity data, and meteorological observation data. The feature construction module is used to construct a continuous numerical electricity price feature sequence based on electricity price policy data, construct a busy season coefficient sequence representing the intensity of busy season based on agricultural activity data and historical charging records of electric vehicles, and extract meteorological feature sequences from meteorological observation data. The feature fusion module is used to align the electricity price feature sequence, agricultural busyness coefficient sequence and meteorological feature sequence with the historical load time series data in terms of time dimension. Based on correlation analysis, it selects key features related to the load and splices the selected key features with the historical load time series data to form a multi-dimensional user behavior feature vector. The model prediction module is used to train a pre-built deep learning load prediction model by inputting multi-dimensional user behavior feature vectors, and then using the deep learning load prediction model to predict the load at future times after the model converges, and output the load prediction results.

[0013] In addition, to achieve the above objectives, this application also proposes an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the rural power distribution network load forecasting method considering electric vehicle user behavior as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a storage medium storing a computer program, which, when executed by a processor, implements the steps of the rural power distribution network load forecasting method considering the behavior of electric vehicle users as described above.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: 1. By acquiring and integrating multi-source data such as electricity pricing policies and agricultural activities, this method fundamentally solves the problem of missing information dimensions caused by relying solely on single historical load data in traditional methods, providing a rich foundation of decision-making information for subsequent forecasting. Furthermore, by constructing discrete electricity pricing policy data as continuous numerical features and building a dynamically calibrable agricultural busyness coefficient based on agricultural activities and charging records, it overcomes the shortcomings of existing technologies that treat electricity prices as categorical variables and agricultural busyness factors as static binary variables. This allows the model to perceive the incentive intensity of continuous changes in electricity prices and the gradual changes and intraday fluctuations in agricultural busyness, thereby accurately quantifying the nonlinear impact of user behavior on load. Based on this, by aligning multi-source features and filtering key features through correlation analysis, a multi-dimensional user behavior feature vector is constructed, ensuring strict synchronization of input data in the time dimension and feature validity, while eliminating redundant and interfering information. Finally, the constructed high-quality multi-dimensional feature vector is input into a deep learning model for training and prediction, enabling the model to learn the inherent complex mapping relationship between user behavior and load changes from the joint data distribution of load, electricity prices, agricultural busyness, and weather, rather than simply fitting historical curves. Therefore, this solution significantly improves the model's ability to interpret load fluctuations and its prediction accuracy, and achieves a significant improvement in the goodness of fit of the prediction results to the actual load. It effectively solves the technical problem of insufficient prediction accuracy caused by ignoring key user behavior factors in the existing technology.

[0016] 2. By breaking down electricity price policy data into three dimensions—base price, price change trend, and relative price level—the model can comprehensively perceive the continuous changing patterns of electricity price signals. It not only identifies the current price level but also captures the guiding effect of price fluctuations on users' decisions to charge earlier or later, and the impact of the current price relative to off-peak electricity prices on users' charging timing choices. Therefore, the model can accurately learn the nonlinear mapping relationship between electricity price changes and load fluctuations, significantly improving the accuracy of load forecasting near electricity price switching periods.

[0017] 3. By constructing a hierarchical agricultural busyness coefficient structure that multiplies a monthly baseline coefficient by an intraday fluctuation coefficient, the model effectively portrays both the monthly agricultural activity cycle and the intraday travel and rest patterns. Furthermore, a data-driven calibration mechanism based on historical charging records is introduced, enabling the agricultural busyness coefficient to dynamically and adaptively adjust according to actual charging behavior characteristics. This overcomes the shortcomings of traditional static assignment methods, which cannot reflect the differences in actual agricultural production intensity across different years and regions. Therefore, the model can accurately capture the temporal distribution differences in user charging behavior during busy and slack seasons, effectively enhancing the fitting ability of load forecasting during seasonal agricultural transitions. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating an embodiment of the rural power distribution network load forecasting method that takes into account the behavior of electric vehicle users in this application; Figure 2 A flowchart illustrating an embodiment of the rural power distribution network load forecasting method that takes into account the behavior of electric vehicle users in this application; Figure 3 A flowchart illustrating an embodiment of the rural power distribution network load forecasting method that takes into account the behavior of electric vehicle users in this application; Figure 4 This is a diagram showing the results of an LSTM load forecasting model that only inputs historical load data, as described in an embodiment of this application. Figure 5 The figure shows the results of the LSTM load forecasting model of the rural power distribution network load forecasting method that considers the behavior of electric vehicle users, according to an embodiment of this application. Figure 6 A schematic diagram of the module structure of a rural power distribution network load forecasting device that takes into account the behavior of electric vehicle users in this application.

[0021] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in 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.

[0023] This application provides a method for load forecasting in rural power distribution networks that considers the behavior of electric vehicle users. Specifically, refer to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the rural power distribution network load forecasting method considering electric vehicle user behavior according to this application. In this embodiment, the rural power distribution network load forecasting method considering electric vehicle user behavior includes steps S10-S40:

[0024] Step S10: Obtain multi-source raw data for the target area. The multi-source raw data includes historical load time series data, electricity price policy data, agricultural activity data, and meteorological observation data.

[0025] It should be noted that electric vehicle user behavior refers to the charging, travel, and parking behaviors of rural residents driven by time-of-use electricity pricing, agricultural activities, and weather conditions. This includes three behavioral patterns: charging shifts in response to electricity prices, charging patterns during busy agricultural seasons, and the influence of temperature, humidity, and rainfall on travel and charging intentions. These are core variables that distinguish rural power distribution network load forecasting from urban load forecasting. Rural power distribution network load forecasting involves short-term time-series load estimation for county and village power distribution networks, with a load resolution of 15 minutes, used for distribution network planning and orderly charging scheduling. Unlike urban distribution networks, it needs to be adapted to the random load fluctuations caused by agricultural activities, electric agricultural machinery, and decentralized private electric vehicle charging by rural residents.

[0026] The multi-source raw data includes historical load time series data, electricity price policy data, agricultural activity data, and meteorological observation data. Historical load time series data refers to the continuous time series of historical active load collected by the distribution network at a fixed time granularity (15 minutes). Electricity price policy data refers to the local power grid's peak / flat / valley time period division and the electricity price data for each time period. Agricultural activity data refers to the agricultural calendar and activity intensity of planting, irrigation, topdressing, and harvesting of crops such as rice, tea, and longan in the county. Meteorological observation data refers to daily / 15-minute average temperature, relative humidity, and rainfall observation records.

[0027] In this embodiment, traditional load forecasting relies solely on historical load sequences, completely ignoring three key behavioral factors unique to rural areas: charging patterns during busy farming seasons, electricity price-driven charging shifts, and weather-related influences on travel intentions. This solution integrates these three behavioral drivers—electricity price, agricultural activities, and weather—to quantify electric vehicle charging behavior from three dimensions: economic, agricultural production, and natural environment. This provides more complete feature information and reduces forecast bias at its root.

[0028] Step S20: Construct a continuous numerical electricity price feature sequence based on electricity price policy data, construct a busy farming coefficient sequence representing the intensity of busy farming based on agricultural activity data and historical charging records of electric vehicles, and extract a meteorological feature sequence from meteorological observation data.

[0029] The continuous numerical electricity price feature sequence abandons the traditional crude treatment of treating peak, flat, and valley prices as discrete categorical variables. It transforms time-of-use electricity prices into continuously calculable numerical time series, including three continuous components: base price, price difference, and price-valley ratio, fully preserving quantitative signals of price change amplitude and relative discount. The agricultural busyness coefficient sequence quantifies the degree of agricultural activity in rural areas and is a dynamic time series parameter that directly reflects the charging time series pattern of electric vehicles, with values ​​ranging from 0 to 1. The coefficient is high during busy agricultural periods (farming during the day and concentrated charging at night) and low during slack agricultural periods (vehicles parked all day and charging time is scattered), breaking through the traditional static binary busy / slack label. The meteorological feature sequence refers to the extraction of average temperature, relative humidity, and rainfall from meteorological data to form a time series vector. Temperature affects electric vehicle battery consumption and residential air conditioning load, while rainfall and humidity reduce residents' willingness to go out, directly changing charging behavior. Daily meteorological data is uniformly expanded to a 15-minute load resolution time series alignment.

[0030] In this embodiment, electricity price reflects users' economic decisions regarding electricity consumption, the agricultural busyness coefficient reflects production and travel constraints, and weather reflects travel and electricity consumption environment constraints; these three types of features are complementary. After fusing these three types of features, the model's goodness of fit is significantly improved compared to single-feature or featureless models, and its explanatory power for load temporal fluctuations is significantly enhanced. Existing user behavior load prediction research focuses on urban areas and does not consider the special load patterns of seasonal agricultural activities in rural areas, electric agricultural machinery, and residents' concentrated charging at night while farming during the day. This application specifically constructs an agricultural busyness coefficient to quantify the impact of agricultural activities, specifically adapted to the county-level rural power distribution network scenario, making the method more targeted for engineering implementation.

[0031] Step S30: Align the electricity price feature sequence, the agricultural busyness coefficient sequence, and the meteorological feature sequence with the historical load time series data in terms of time dimension. Based on correlation analysis, select key features related to the load, and then concatenate the selected key features with the historical load time series data to form a multi-dimensional user behavior feature vector.

[0032] The time index and time resolution (15 minutes) of all features and load data are unified to ensure a one-to-one correspondence between load, electricity price, agricultural busy season, and meteorological data at the same time step, eliminating feature failure caused by time series misalignment and providing a foundation for feature splicing. Correlation analysis employs the Spearman correlation coefficient method to quantify the correlation strength between each external feature and the distribution network load, filtering out low-correlation invalid features, reducing the dimensionality of the model input, and minimizing redundant noise. Key features refer to those with correlation coefficients higher than a preset threshold after correlation screening, possessing significant explanatory power for load fluctuations. In this embodiment, three types of key features are screened: electricity price features, agricultural busy season coefficient, and average temperature.

[0033] Step S40: Input the multidimensional user behavior feature vector into the pre-built deep learning load prediction model for training, and use the deep learning load prediction model to predict the load at future times after the model converges, and output the load prediction result.

[0034] A multidimensional user behavior feature vector refers to a multidimensional time-series input vector formed by concatenating historical load values ​​with all filtered key behavioral features at a single time step. Each time step simultaneously includes four types of information: grid base load, user electricity price response behavior, agricultural travel behavior, and weather-constrained behavior. Deep learning load forecasting models, specifically Long Short-Term Memory (LSTM) networks, are adapted to the time-dependent characteristics of load time-series data, overcoming the limitation of traditional machine learning in capturing long-term load patterns.

[0035] Model training involves dividing the normalized multidimensional feature vectors into a training set, constructing time-series samples using a sliding window, iteratively updating network weights and bias parameters, and using an optimizer and gradient pruning to stabilize the training process until the loss converges. For the converged model, inputting test time-series samples, outputs normalized load prediction values, which are then denormalized to restore the actual power values, forming a continuous load time-series prediction curve and numerical results.

[0036] In this embodiment, all features are standardized to a 15-minute time resolution and time index alignment to avoid feature failure due to temporal misalignment. Low-correlation redundant meteorological indicators are eliminated through correlation screening, reducing the model input dimension and accelerating training convergence without losing effective information, thus reducing computational consumption. Since load and charging behavior exhibit strong temporal correlations, traditional statistical methods and shallow machine learning cannot capture multi-day, multi-period temporal variation patterns. LSTM relies on a gating mechanism to memorize historical temporal information, adapting to the random and time-varying characteristics of rural electric vehicle charging loads, and exhibits stronger generalization performance compared to traditional prediction models.

[0037] In one implementation, reference is made to Figure 2 Step S20 includes: Step S201: Determine the electricity price period for each time point based on the electricity price policy data, and use the electricity price value corresponding to each time point as the basic electricity price component.

[0038] Match each 15-minute time index to the local peak / slow / valley time period and directly assign the corresponding time period electricity price as the base electricity price component.

[0039] Step S202: Based on the basic electricity price component, extract the electricity price difference between adjacent time points as the electricity price change trend component.

[0040] Let the time resolution of the load sequence be... The total time step is The time index is Assume that the time-of-use electricity pricing policy data includes... The time period, the first The electricity price for each time period is Based on local time-of-use pricing policies, discrete time periods are transformed into continuous numerical pricing characteristics.

[0041] This represents the price difference between adjacent time periods. It is defined as the difference between the current electricity price and the price at the previous time point. This indicates that electricity prices are rising. This indicates that electricity prices are decreasing. This component depicts the direction and rate of change in electricity prices, and is crucial for understanding users' intertemporal shifts in electricity usage, such as bringing forward or postponing it.

[0042] Step S203: Extract the ratio between the current electricity price and the lowest electricity price of the day as the relative level component of the electricity price.

[0043] To enable the model to not only perceive the absolute value of electricity prices but also capture price trends and relative affordability, two key components are further extracted from the basic electricity price series: the price difference between adjacent time periods and the ratio of the current electricity price to the off-peak electricity price of the day. For the current electricity price Electricity price during off-peak hours of the day The ratio, , This is the lowest electricity price during all time periods of the day (if there are multiple off-peak periods, take the minimum value). This indicates that the current electricity price is more expensive than off-peak electricity, reflecting the opportunity cost of the current electricity price.

[0044] Step S204: Combine the base electricity price component, the electricity price change trend component, and the electricity price relative level component to form a continuous numerical electricity price feature sequence.

[0045] The base electricity price component, the electricity price trend component, and the relative electricity price level component are combined to form the electricity price feature vector at a continuous numerical time t. :

[0046] The electricity price feature vector at each time step forms a continuous electricity price feature sequence with the same length and resolution as the load time series, fully preserving the three price signals of absolute price, changing trend, and relative discount, thereby improving the model's ability to learn from electricity price response behavior.

[0047] In this embodiment, traditional methods use peak, flat, and valley as classification labels, losing information on the specific amount of electricity price and the size of the price difference, making it impossible for the model to distinguish between high-price peak periods and low-price peak periods. This solution uses continuous numerical modeling, allowing the model to accurately perceive differences in electricity prices and fully learn users' electricity price response behavior. When electricity prices rise, users will charge earlier, and when electricity prices fall, users will delay charging. Relying solely on static electricity price values ​​cannot identify this dynamic behavior. The price difference between adjacent time periods can intuitively reflect the magnitude and direction of electricity price fluctuations, quantifying users' motivation for shifting electricity consumption across time periods and improving the accuracy of peak-valley charging load prediction. Rural users generally prefer charging at lower prices during off-peak hours. The ratio of the current electricity price to off-peak electricity can intuitively reflect the cost of electricity; the larger the ratio, the lower the willingness to charge. This component accurately models the core load characteristics of concentrated charging of rural electric vehicles during off-peak hours at night, significantly reducing nighttime load prediction errors. The base electricity price reflects absolute cost, the price difference reflects price change trends, and the off-peak electricity ratio reflects the relative degree of discount. The three types of component information are non-redundant, and their combination completely covers all dimensions of electricity price's influence on users' charging decisions, resulting in a significant improvement in prediction accuracy compared to single electricity price features.

[0048] In one implementation, reference is made to Figure 3 Step S20 also includes: Step S205: Determine the monthly baseline agricultural busyness coefficient for each month based on agricultural activity data.

[0049] Based on the agricultural calendar of planting, irrigation and harvesting of crops such as rice, tea and longan in the county, the monthly busy and slack seasons are divided. The baseline coefficient for busy months is higher (in the range of 0 to 1), while that for slack months is lower, thus depicting the changes in agricultural intensity throughout the year on a monthly scale.

[0050] Step S206: Divide the daytime into periods based on the travel and rest patterns of rural residents, and assign a daytime fluctuation coefficient to each period.

[0051] The system is divided into time periods: early morning, morning, morning busy farming season, lunch break, afternoon busy farming season, evening, and night. Different time periods are assigned a coefficient of 0 to 1 to reflect the daily charging patterns: the coefficient is high during busy farming seasons (farming outside, no charging) and high at night (charging upon returning home), thus depicting the daily load fluctuation pattern.

[0052] Step S207: Multiply the monthly baseline agricultural busyness coefficient by the intraday fluctuation coefficient of the corresponding date to obtain the stratified agricultural busyness coefficient, which serves as the base value of the agricultural busyness coefficient series.

[0053] Define the monthly baseline agricultural busyness coefficient ( Intraday volatility coefficient ( ), then time The corresponding stratified busy farming coefficients are:

[0054] In the formula, For a time-indexed segment, the hierarchical agricultural busyness coefficient is... The agricultural busyness coefficient for each location, For a moment The corresponding month index, For a moment The corresponding hourly time period index.

[0055] Step S208: Obtain the historical charging records of electric vehicles, extract the charging behavior features from the historical charging records, perform data-driven dynamic calibration of the hierarchical agricultural busyness coefficient based on the charging behavior features, and generate the calibrated agricultural busyness coefficient sequence.

[0056] Traditional agricultural busyness coefficients rely solely on static assignments from agricultural calendars. This application uses real EV charging data to reverse-correct the coefficients. Agricultural calendars serve as prior constraints, while charging records provide feedback on real-world behavior. Through data-driven calibration, static empirical parameters are transformed into learnable dynamic parameters, improving the matching degree between agricultural busyness characteristics and real charging loads.

[0057] In this embodiment, due to the annual monthly cycle (high load during sowing and harvesting seasons) and daily cycle (no charging during the day and concentrated charging at night), the traditional binary label of busy / idle cannot distinguish the intensity of busy seasons or the differences in daily time periods. The stratified coefficient takes into account both monthly agricultural intensity and daily travel patterns, resulting in a more granular characterization of busy season load. During the same agricultural season, different villages have different electric vehicle ownership and travel habits. By reverse-calibrating the coefficient through real charging records, static empirical parameters are upgraded to dynamic, learnable parameters adapted to local users, ensuring that busy season characteristics align with local actual load patterns. The agricultural calendar provides objective prior constraints on agricultural production, preventing complete distortion of the coefficient when charging data is missing; charging records provide real user behavior observation data, correcting the idealization bias of the agricultural calendar. The fusion of these two data sources significantly improves the reliability of the busy season coefficient.

[0058] In one implementation, step S208 includes: The charging frequency, average charging amount, concentration of charging time period, and average charging duration per unit time are extracted from historical charging records as charging behavior features. The charging behavior features are input into a preset mapping function, and the output is an initial estimate of the busy farming intensity derived from the charging behavior. The initial estimate of the busy farming intensity is fused with the prior busy farming intensity determined based on agricultural activity data to obtain the calibrated busy farming intensity, so as to update the hierarchical busy farming coefficient.

[0059] Traditional methods rely on static assignments of agricultural calendar dates (sowing, irrigation, harvesting, etc.), which cannot fully capture the correlation between agricultural production and electric vehicle charging. This solution uses historical EV charging records to dynamically infer the intensity of agricultural activity. The core logic is that during busy farming seasons (such as the sowing and harvesting of early and late rice), rural residents' main activities shift to fieldwork, and their travel behavior exhibits a pattern of "high-frequency short-distance travel during the day and concentrated parking at night." After nighttime work, vehicles return to their residences, and charging activity is concentrated at night, forming a temporal distribution of "no charging during the day and concentrated charging at night." During the off-season, user travel demand decreases significantly, vehicles are mostly parked for extended periods, and charging time is more flexible. Therefore, the EV charging behavior feature vector at time t is defined as:

[0060]

[0061] in This refers to the charging frequency (the number of times a device is charged per unit of time). Average charging capacity (kWh). Concentration of charging periods Average charging duration.

[0062] Through mapping function The initial estimate for mapping charging behavior characteristics to the intensity of busy farming season is as follows:

[0063] in , , , For learnable weights, This is a bias term.

[0064] At the same time, the agricultural calendar is introduced as a priori constraint to define the priori intensity of agricultural busyness:

[0065] in The number of key agricultural activities in the agricultural calendar. Let K be the intensity coefficient of the activity. As the center of the event time, This is the time span parameter for the activity.

[0066] By back-calculating the EV charging data with the prior agricultural calendar, a Bayesian fusion was performed to obtain the calibrated agricultural busyness intensity.

[0067] in, For the uncertainty of the a priori agricultural calendar, Uncertainty derived from EV data.

[0068] In this implementation, charging during the busy farming season is concentrated at night, with high concentration and extremely low daytime charging frequency; during the off-season, charging is dispersed and evenly distributed throughout the day. These four characteristics differentiate the busy farming season from multiple perspectives—charging quantity, charging scale, time distribution, and duration—providing sufficient quantitative evidence for inferring the intensity of the busy season. The mapping function has built-in trainable weights and biases, which can adaptively adjust the contribution weights of each charging characteristic to the intensity of the busy season based on local charging data, eliminating the need for manual coefficient setting. It is highly versatile and adaptable to rural areas with different planting industries and varying electric vehicle ownership. During fusion, the uncertainty of each type of data is weighted, mitigating the influence of data sources with larger errors, resulting in a more accurate calibrated busy season coefficient. As agricultural activities and electric vehicle ownership change annually, charging behavior patterns will slowly change. This calibration process can periodically import new charging records to update the busy season coefficient, allowing the model to adapt to the dynamically changing user travel and charging behavior in rural areas over the long term, without requiring manual recalibration of the agricultural coefficient.

[0069] In one implementation, step S30 further includes: using the Spearman correlation coefficient method to calculate the correlation measure between the electricity price feature sequence, the agricultural busyness coefficient sequence, the meteorological feature sequence and the historical load time series data respectively; identifying features with correlation measure values ​​greater than a preset threshold as key features; aligning the key features with the historical load time series data in the time dimension, and splicing them by time step, so that each time step contains the load value and its corresponding key feature value, forming a multi-dimensional user behavior feature vector.

[0070] It should be noted that the meteorological feature sequence construction process involves extracting average temperature (°C), relative humidity (%), and rainfall (mm) from meteorological observation data. Considering that county-level meteorological data is usually provided on a daily basis, it can be extended to the same temporal resolution as the load (i.e., repeating the meteorological values ​​of that day at all time points). Let... The average temperature, relative humidity, and rainfall at each time point were respectively , , Then the meteorological feature vector sequence is:

[0071]

[0072] In the formula, It is a sequence of meteorological feature vectors. This is a sequence of meteorological feature vectors for a time-indexed segment. For 15-minute meteorological observations, direct alignment is used. Temperature reflects battery performance and air conditioning energy consumption, while humidity and rainfall reflect travel intentions.

[0073] The electricity price characteristic sequence, agricultural busyness coefficient sequence, and meteorological characteristic sequence obtained from the above embodiments are strictly aligned with the original historical load sequence in the time dimension. The Spearman rank correlation coefficient method is used to screen features with high linear correlation to the load sequence. The core calculation formula of the Spearman correlation coefficient method is:

[0074] Where n is the sample size and d is the rank difference.

[0075] The features filtered by the Spearman correlation coefficient method are concatenated into a multidimensional user behavior feature vector. Each time step of the model input vector should include four-dimensional features: historical load, electricity price, agricultural busyness factor, and average temperature. The final model input vector sequence is as follows:

[0076] In the formula, For the model input vector, Let the electricity price vector be... This is a vector of agricultural busyness coefficients. The average temperature vector. This is the load vector.

[0077] In this implementation, Spearman calculates correlation based on rank, which is not constrained by the linearity of variables and adapts to nonlinear time-series power load scenarios, accurately identifying the effective features that truly affect load fluctuations. Only highly correlated factors such as electricity price, agricultural busyness factor, and average temperature are retained. After simplifying the input feature dimensions, the number of model parameters is reduced, effectively suppressing overfitting problems under small-sample time-series training and improving the model's generalization ability. After unifying the time index and aligning with 15-minute granularity, each time step feature is matched one-to-one with the load, ensuring that the physical logic of the input samples is real and effective. All samples use feature vectors with a unified dimension and structure, eliminating the need for additional segmented data processing. They can be directly fed into a sliding window to construct time-series samples, simplifying the data preprocessing process and reducing the complexity of engineering code development.

[0078] In one implementation, step S40 includes: The multidimensional user behavior feature vectors are normalized to eliminate the differences in dimensions between features. A sliding window method is used to construct a training sample set from the normalized multidimensional user behavior feature vectors. The training sample set is then input into a long short-term memory network model, which consists of an input layer, an LSTM hidden layer, a Dropout layer, a fully connected layer, and an output layer connected in sequence. An optimizer combined with a gradient pruning strategy is used to optimize the parameters of the long short-term memory network model, and the model training is completed when the convergence condition is met.

[0079] The deep learning load prediction model is a Long Short-Term Memory (LSTM) network model. The constructed multi-dimensional user behavior feature vector is normalized using Min-Max to eliminate differences in units and magnitudes between different variables before being input into the LSTM network. The MinMax normalization calculation formula is as follows:

[0080]

[0081] In the formula, The output result after normalization The original characteristics or the true values ​​of the load. , These are the global maximum and minimum values ​​of the variables in this group, respectively.

[0082] LSTM uses memory cells and three gating mechanisms: an input gate, a forget gate, and an output gate. Each gate uses a sigmoid function to control the flow of information, and the memory cells are updated using a tanh function. The formula for calculating the forget gate is:

[0083]

[0084] In the formula, Output for the forget gate. It is the Sigmoid activation function. , These correspond to the input weights and loop weights of the forget gate, respectively. Input the feature vector at the current time. The state was hidden in the previous moment. This is a bias term.

[0085] The formula for calculating the input gate is:

[0086] In the formula, The output of the input gate, , These correspond to the weights of the input gate and the recurrent weights, respectively. This is the bias term. Simultaneously, candidate memory vectors are generated using the Tanh activation function:

[0087]

[0088] In the formula, The hyperbolic tangent activation function is used. , These correspond to the weights of the input gate and the recurrent weights, respectively. This is a bias term.

[0089] The forget gate is multiplied by the memory unit from the previous time step, discarding unnecessary information. The input gate is then multiplied by the candidate memory unit to obtain the new information to be added. The two are added together to obtain the new memory unit for the current time step, completing the overall update of the cell state. The update formula is:

[0090]

[0091] In the formula, For the current memory unit, Given the cell state at the previous moment, this step enables the fusion and iteration of historical memory and new information.

[0092] Output gate The expression is:

[0093] In the formula, , These correspond to the weights of the input gate and the recurrent weights, respectively. This is a bias term.

[0094] Then, the memory cells are compressed to the range (-1, 1) using the tanh activation function, and then multiplied one by one with the output gate elements to obtain the final hidden state. Final Hidden State for:

[0095]

[0096] A sliding window is used to construct training samples (predicting the next point using the past 24 time points), followed by model training. The model training is set to 50 epochs, with 64 iterations per epoch and a maximum of 3200 iterations. The Adam optimizer is used for gradient updates, and the gradient norm is constrained when updating parameters.

[0097] In the formula, The gradient vector of the model parameters. The Euclidean norm of the gradient. The threshold for gradient clipping is set to 1.0.

[0098] Adam optimizer parameter update rules:

[0099] In the formula, These are the parameters updated at step t+1. Let be the model parameters at step t. The global learning rate, , These are the first and second moment estimates after bias correction, respectively.

[0100] A "Model Convergence" checkpoint is set during training. If the model loss does not stabilize or the preset maximum number of iterations is not reached, it is considered non-converged. The model configuration is optimized by adjusting hyperparameters, and the parameter initialization step is returned for retraining until the model converges. If the model has converged, the subsequent prediction and evaluation phases are initiated to ensure that the model performance reaches its optimal state.

[0101] After the model converges, recursive prediction is performed. The trained LSTM model is used to perform time-series predictions on the test set data, outputting normalized load prediction results. Then, an "inverse normalization" operation is used to restore the predicted values ​​to their original dimensions, obtaining the actual load prediction sequence. Finally, the predicted load curve and specific values ​​are output, completing the core prediction task.

[0102] In this embodiment, the numerical magnitudes of load (kW), electricity price (yuan / kWh), temperature (°C), and agricultural busyness coefficient (dimensionless) differ significantly, and direct training would lead to network weight imbalance and gradient explosion. Normalization maps all features to the [0,1] interval, balancing the impact of each feature on the model and accelerating convergence. The load at the next moment is predicted using 24 historical time steps, with the window sliding step-by-step to extract samples, fully utilizing the temporal correlation information before and after the load. This generates several times the number of training samples under the same original data, ensuring model training effectiveness even in small dataset scenarios. The LSTM three-gating mechanism can memorize long-term load change patterns over multiple days and time periods, addressing the shortcomings of traditional fully connected networks in handling time-series data; the Dropout layer randomly shields some neurons, reducing the risk of overfitting in time-series models and minimizing prediction errors in unfamiliar scenarios. LSTM deep time-series networks are prone to gradient explosion leading to training crashes; Adam adaptively adjusts the learning rate to accelerate convergence; gradient pruning constrains the maximum norm of the gradient, limiting the infinite increase of the gradient from a numerical perspective and ensuring stability throughout the training of large-scale time-series data. The system determines whether to terminate training based on the stability of the loss, avoiding underfitting (early termination) and overfitting (excessive iteration). If the training fails to converge, the system automatically adjusts the hyperparameters for retraining, eliminating the need for manual monitoring of the training process. This high degree of automation results in the output of the optimal weight model.

[0103] In one implementation, step S40 is followed by: Acquire multi-source raw data under various test scenarios, including at least different seasonal scenarios, different date types, and different electricity price periods; input the multi-source raw data under each test scenario into the trained deep learning load prediction model to obtain the load prediction results corresponding to each test scenario; evaluate the generalization ability and robustness of the deep learning load prediction model based on the load prediction results corresponding to each test scenario and the actual load data under the corresponding scenario.

[0104] This example uses one year's worth of data from a specific county, with a time resolution of 15 minutes. It includes the following steps:

[0105] Step 1: Obtain the following data for the target area within the historical time period: Electric vehicle charging record dataset: Includes fields such as charging start time, charging end time, charging amount (kWh), and charging power for each charging pile / station. The electric vehicles include electric agricultural machinery, electric transport vehicles, and electric passenger vehicles for rural residents. The charging record data is aggregated according to a preset time granularity (preferably 1 hour) to obtain the charging frequency, average charging amount, charging period concentration, and average charging duration within each time window.

[0106] Agricultural calendar prior knowledge base: includes the planting types of major crops in the target area, the time nodes of key agricultural activities such as sowing time windows, irrigation time windows, topdressing time windows, and harvesting time windows, as well as their typical durations.

[0107] Historical load data: Historical load time-series data of the target area's distribution network.

[0108] Meteorological data: including temperature, precipitation, sunshine duration, etc.

[0109] Step 2: Construct a feature vector based on the data obtained in Step 1. When constructing the electricity price feature vector, determine the electricity price period for each 15-minute time point based on its time. When constructing the meteorological features, for the forecast day, the average temperature is 28.5℃, the relative humidity is 75%, and the rainfall is 0mm. Assign the same temperature, humidity, and rainfall values ​​to all 96 15-minute points of that day. Finally, align the above features with the historical load according to the time points to form a four-dimensional vector.

[0110] Step 3: Strictly align the electricity price feature sequence, agricultural busyness coefficient sequence, and meteorological feature sequence obtained in Step 2 with the original historical load sequence in the time dimension, and use Spearman correlation analysis. The results show that the average temperature, electricity price coefficient, and agricultural busyness coefficient have the highest correlation with the load, and are therefore determined as the final input features.

[0111] Step 4: Input the input vector that integrates behavioral features and load data into the LSTM model (64 neurons, dropout 0.2, Adam optimizer, sliding window 24). On the test set, the results are R²=0.9302, RMSE=6.33kW, and MAPE=3.24%. Compared with the model that only uses historical load data, the accuracy is significantly improved, proving that the user behavior features constructed in this application can effectively capture the impact of busy farming seasons and electricity prices on load.

[0112] Step 5: Validate the model performance in different seasons, date types, and time periods to ensure the method has engineering applicability.

[0113] In summary, the rural power distribution network load forecasting method proposed in this application, which considers the behavior of electric vehicle users, has an LSTM load forecasting model that integrates three-dimensional features of electricity price, average temperature, and agricultural busyness factor. The evaluation index of this model is better than that of the LSTM load forecasting model that only inputs historical load. This method can effectively improve the prediction accuracy of the LSTM load forecasting model for power distribution networks in county-level scenarios and provide a theoretical basis for rural power distribution network planning.

[0114] In the embodiments of this application, to verify the effectiveness of this application, the results of the rural power distribution network load prediction model are evaluated based on the following ablation experiments: 1) Method 1: Do not integrate electric vehicle user behavior characteristics, only input historical load data; 2) Method 2: The rural power distribution network load forecasting method of this application that considers the behavioral characteristics of electric vehicle users integrates different behavioral characteristics of historical load data, including electricity price characteristics, meteorological characteristics and busy farming season characteristics; The LSTM load forecasting model integrating different features was evaluated using four evaluation indicators: R², RMSE, MAE, and MAPE. The results are shown in Table 1.

[0115] Table 1 Evaluation Indicators of LSTM Load Forecasting Models Integrating Different Features

[0116] As shown in Table 1: 1) Compared with Method 1, the LSTM load forecasting model with multi-behavioral feature fusion proposed in Method 2 of this application has a stronger explanatory power for load time series changes. The fusion of multiple source features such as electricity price, average temperature, and busy farming season has a complementary effect, improving the prediction accuracy by 9.58% and achieving a higher goodness of fit to the true value. 2) Among the behavioral features proposed in this application, the agricultural busy season feature is an important behavioral driver variable of county-level charging load. After integrating the agricultural busy season feature, the model prediction accuracy is improved by 7.83%. The hierarchical agricultural busy season coefficient of this application can more finely characterize the time-varying impact of agricultural busy season on load and has a good prediction effect.

[0117] Therefore, the rural power distribution network load forecasting method that considers the behavioral characteristics of electric vehicle users proposed in this application can effectively improve the load forecasting accuracy of rural power distribution networks, which is of great significance for the safe and stable operation of the power system.

[0118] This scheme transforms the discrete time periods in the time-of-use pricing policy into continuous numerical features, preserving the strength information of the price signal, which is beneficial for the model to learn the negative correlation between electricity price and load. For the first time, it uses historical EV charging records and agricultural calendars of major crops to determine the agricultural busyness coefficient, introducing a data-driven calibration mechanism. This upgrades the agricultural busyness coefficient from a static empirical value to a learnable dynamic behavioral parameter, significantly improving the accuracy of characterizing the impact of agricultural busyness factors on load. This constructs a multi-dimensional user behavior feature vector, including historical load, electricity price, agricultural busyness coefficient, and meteorological features, comprehensively representing electric vehicle user behavior from economic, social, and natural dimensions, providing high-quality input for deep learning load forecasting. Experimental results show that the LSTM load forecasting model using the feature construction method of this application achieves an R² of 0.9302, a 9.58% improvement compared to the baseline model using only historical load, a 35.18% reduction in RMSE, and a 2.34 percentage point reduction in MAPE, providing theoretical support for rural distribution network planning and orderly charging strategies.

[0119] Based on the same inventive concept, such as Figure 6 As shown, this application also provides a rural power distribution network load forecasting device that considers the behavior of electric vehicle users. The rural power distribution network load forecasting device that considers the behavior of electric vehicle users includes: Data acquisition module 10 is used to acquire multi-source raw data of the target area, including historical load time series data, electricity price policy data, agricultural activity data and meteorological observation data; The feature construction module 20 is used to construct a continuous numerical electricity price feature sequence based on electricity price policy data, construct a busy farming coefficient sequence representing the intensity of busy farming based on agricultural activity data and historical charging records of electric vehicles, and extract meteorological feature sequences from meteorological observation data. The feature fusion module 30 is used to align the electricity price feature sequence, the agricultural busyness coefficient sequence and the meteorological feature sequence with the historical load time series data in the time dimension, respectively, to screen out key features related to the load based on correlation analysis, and to splice the screened key features with the historical load time series data to form a multi-dimensional user behavior feature vector. The model prediction module 40 is used to input multi-dimensional user behavior feature vectors into a pre-built deep learning load prediction model for training, and after the model converges, use the deep learning load prediction model to predict the load at future times and output the load prediction results.

[0120] It should be noted that the rural power distribution network load forecasting device considering electric vehicle user behavior provided in this application embodiment and the rural power distribution network load forecasting method considering electric vehicle user behavior provided in this application embodiment are based on the same application concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned rural power distribution network load forecasting method considering electric vehicle user behavior, and the repeated parts will not be described again.

[0121] In some embodiments, an electronic device provided in this application includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the above-described method for predicting rural power distribution network loads that takes into account the behavior of electric vehicle users.

[0122] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0123] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0124] This application also provides a storage medium storing a computer program that, when executed by a processor, implements the aforementioned rural power distribution network load forecasting method considering electric vehicle user behavior. This storage medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not integrated into that device / apparatus / system. The storage medium carries one or more programs, which, when executed, implement the method as described in the embodiments or implementations of this application.

[0125] According to embodiments of this application, the storage medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. A computer-readable signal medium may also be any storage medium other than a computer-readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the storage medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0126] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0127] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for load forecasting in rural power distribution networks that considers the behavior of electric vehicle users, characterized in that, The method includes: Acquire multi-source raw data for the target area, including historical load time-series data, electricity price policy data, agricultural activity data, and meteorological observation data; A continuous numerical electricity price feature sequence is constructed based on the electricity price policy data; a busy farming coefficient sequence is constructed based on the agricultural activity data and the historical charging records of electric vehicles; and a meteorological feature sequence is extracted from the meteorological observation data. The electricity price feature sequence, agricultural busyness coefficient sequence, and meteorological feature sequence are aligned with the historical load time series data in terms of time dimension. Based on correlation analysis, key features related to the load are selected, and the selected key features are concatenated with the historical load time series data to form a multi-dimensional user behavior feature vector. The multidimensional user behavior feature vector is input into a pre-built deep learning load prediction model for training. After the model converges, the deep learning load prediction model is used to predict the load at future times, and the load prediction result is output.

2. The method as described in claim 1, characterized in that, The steps for constructing a continuous numerical electricity price feature sequence include: Based on the electricity price policy data, determine the electricity price period to which each time point belongs, and use the electricity price value corresponding to each time point as the basic electricity price component; Based on the basic electricity price component, the electricity price difference between adjacent time points is extracted as the electricity price change trend component; Extract the ratio between the current electricity price and the lowest electricity price of the day as the relative level component of the electricity price; The basic electricity price component, the electricity price change trend component, and the electricity price relative level component are combined to form the continuous numerical electricity price feature sequence.

3. The method as described in claim 1, characterized in that, The steps for constructing the agricultural busyness coefficient sequence characterizing the intensity of agricultural busyness include: The monthly baseline agricultural activity coefficient for each month is determined based on the aforementioned agricultural activity data; The daily time periods are divided according to the travel and work patterns of rural residents, and a daily fluctuation coefficient is assigned to each daily time period. The monthly baseline agricultural busyness coefficient is multiplied by the intraday fluctuation coefficient of the corresponding date to obtain the stratified agricultural busyness coefficient, which serves as the base value of the agricultural busyness coefficient sequence. The historical charging records of electric vehicles are obtained, the charging behavior features in the historical charging records are extracted, and the hierarchical agricultural busyness coefficient is dynamically calibrated based on the charging behavior features to generate a calibrated agricultural busyness coefficient sequence.

4. The method as described in claim 3, characterized in that, The step of performing data-driven dynamic calibration of the stratified agricultural busyness coefficient based on the charging behavior characteristics includes: The charging frequency, average charging amount, concentration of charging time period, and average charging duration are extracted from the historical charging records and used as the charging behavior characteristics. The charging behavior characteristics are input into a preset mapping function, and the output is an initial estimate of the busy farming intensity derived from the charging behavior. The initial estimate of the busy farming season intensity is fused with the prior busy farming season intensity determined based on the agricultural activity data to obtain the calibrated busy farming season intensity, thereby updating the stratified busy farming season coefficient.

5. The method as described in claim 1, characterized in that, The step of concatenating the selected key features with the historical load time-series data to form a multidimensional user behavior feature vector includes: The Spearman correlation coefficient method was used to calculate the correlation measure between the electricity price characteristic sequence, the agricultural busyness coefficient sequence, the meteorological characteristic sequence and the historical load time series data, respectively. Features whose correlation metric values ​​are greater than a preset threshold are identified as key features; The key features are aligned with the historical load time series data in the time dimension and concatenated by time step so that each time step contains the load value and its corresponding key feature value, thus forming the multidimensional user behavior feature vector.

6. The method as described in claim 1, characterized in that, The deep learning load prediction model is a long short-term memory network model. The step of inputting the multi-dimensional user behavior feature vector into the pre-constructed deep learning load prediction model for training includes: The multidimensional user behavior feature vector is normalized to eliminate the dimensional differences between the features; A training sample set is constructed from the normalized multidimensional user behavior feature vector using a sliding window approach; The training sample set is input into the long short-term memory network model, which includes an input layer, an LSTM hidden layer, a Dropout layer, a fully connected layer, and an output layer connected in sequence. The parameters of the long short-term memory network model are optimized using an optimizer combined with a gradient pruning strategy, and the model training is completed when the convergence condition is met.

7. The method as described in claim 1, characterized in that, Following the step of outputting the load prediction results, the method further includes: Acquire multi-source raw data under various test scenarios, including at least different seasonal scenarios, different date type scenarios, and different electricity price time period scenarios; The multi-source raw data from each test scenario are input into the trained deep learning load prediction model to obtain the load prediction results for each test scenario. Based on the load prediction results for each test scenario and the actual load data for that scenario, the generalization ability and robustness of the deep learning load prediction model are evaluated.

8. A rural power distribution network load forecasting device considering the behavior of electric vehicle users, characterized in that, The rural power distribution network load forecasting device that considers electric vehicle user behavior includes: The data acquisition module is used to acquire multi-source raw data of the target area, including historical load time series data, electricity price policy data, agricultural activity data, and meteorological observation data. The feature construction module is used to construct a continuous numerical electricity price feature sequence based on the electricity price policy data, construct a busy farming coefficient sequence representing the intensity of busy farming based on the agricultural activity data and the historical charging records of electric vehicles, and extract a meteorological feature sequence from the meteorological observation data. The feature fusion module is used to align the electricity price feature sequence, agricultural busyness coefficient sequence and meteorological feature sequence with the historical load time series data in terms of time dimension, filter out key features related to load based on correlation analysis, and splice the selected key features with the historical load time series data to form a multi-dimensional user behavior feature vector. The model prediction module is used to input the multidimensional user behavior feature vector into a pre-built deep learning load prediction model for training, and after the model converges, use the deep learning load prediction model to predict the load at future times and output the load prediction result.

9. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the rural power distribution network load forecasting method considering electric vehicle user behavior as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the steps of the rural power distribution network load forecasting method considering electric vehicle user behavior as described in any one of claims 1 to 7.

Images