Interconnection line load prediction method and device, electronic equipment and storage medium

Abstract

This application relates to the field of power grid dispatching and provides a method, device, electronic equipment, and storage medium for tie-line load forecasting. The method includes: feature filtering and reconstruction of each historical multi-dimensional data point to obtain corresponding first candidate training samples; selecting multiple target training samples from multiple first candidate training samples; using the multiple target training samples to train an initial tie-line load forecasting model to obtain a trained tie-line load forecasting model; acquiring the multi-dimensional data to be predicted for the date to be predicted and inputting it into the trained tie-line load forecasting model, and outputting the tie-line load forecasting result for the date to be predicted. This application can systematically integrate the core dynamic influencing factors that affect and determine the power output of tie lines, replacing subjective experience judgment with a quantitative analysis mechanism, significantly improving the accuracy and reliability of the forecasting model, thereby meeting the actual needs of high-precision tie-line load forecasting in the construction of a unified inter-provincial power market.

Description

Technical Field

[0001] This application relates to the field of power grid dispatching, and in particular to a tie-line load forecasting method, apparatus, electronic device, and storage medium. Background Technology

[0002] In the process of inter-provincial and regional power trading and the construction of a unified market, tie-line load is a core variable that determines the formulation of medium- and long-term trading prices, the implementation of congestion management, and the optimization and control of inter-regional power flow. Achieving high-precision and interpretable medium- and long-term forecasts of tie-line load is a key technical prerequisite for market participants to conduct inter-regional trading decision-making, transmission right valuation, and congestion risk avoidance.

[0003] Currently, tie-line load forecasting for electricity market scenarios, especially high-precision and interpretable forecasting tasks serving medium- and long-term price forecasting, mainly employs extrapolation methods based on historical average statistics, agreement execution, and adjustments made based on human experience. In current market practice, the output of such forecasts largely depends on fixed power transmission and reception agreements, historical average power transmission data for the same period, or adjustments made by dispatchers based on subjective experience and limited boundary information.

[0004] However, the aforementioned traditional forecasting methods have significant drawbacks: their forecasting logic is highly simplified and static, and they are severely inadequate in responding to dynamic changes in the electricity market, completely ignoring the core dynamic factors that influence and determine the size of tie-line power; at the same time, these methods lack data-driven quantitative analysis mechanisms as support, making it difficult for the accuracy and reliability of the forecast results to meet the actual needs of building a unified cross-provincial electricity market. Summary of the Invention

[0005] In view of this, embodiments of this application provide a tie-line load forecasting method, apparatus, electronic device, and storage medium to address the problems in the prior art where the forecasting logic is highly simplified and static, resulting in insufficient response to dynamic changes in the electricity market and a lack of data-driven quantitative analysis mechanisms, which makes it difficult to meet the actual needs of the construction of a unified cross-provincial electricity market in terms of accuracy and reliability.

[0006] A first aspect of this application provides a tie-line load forecasting method, comprising: Starting from the date to be predicted, multiple historical multi-dimensional data points are collected backward; each historical multi-dimensional data point includes at least the historical boundary feature information and historical meteorological feature information of its corresponding historical date; For each historical multi-dimensional data point, feature filtering and reconstruction are performed on the historical multi-dimensional data to obtain the corresponding first candidate training sample; the total number of first candidate training samples is equal to the total number of historical multi-dimensional data points. Multiple target training samples are selected from multiple first candidate training samples; wherein the total number of target training samples is less than or equal to the total number of first candidate training samples. The initial tie-line load prediction model is trained using multiple target training samples until the preset convergence condition is met, thus obtaining the trained tie-line load prediction model. The system acquires the multi-dimensional data corresponding to the date to be predicted and inputs it into the trained tie-line load prediction model. The system outputs the tie-line load prediction results for the date to be predicted. The multi-dimensional data to be predicted includes boundary feature information and meteorological feature information corresponding to the date to be predicted.

[0007] A second aspect of this application provides a tie-line load prediction device, comprising: The data acquisition module is configured to collect multiple historical multi-dimensional data points backward from the date to be predicted; each historical multi-dimensional data point includes at least the historical boundary feature information and historical meteorological feature information of its corresponding historical date. The first filtering module is configured to perform feature filtering and reconstruction on each historical multi-dimensional data point to obtain the corresponding first candidate training sample; the total number of first candidate training samples is equal to the total number of historical multi-dimensional data points. The second filtering module is configured to filter out multiple target training samples from multiple first candidate training samples; wherein the total number of target training samples is less than or equal to the total number of first candidate training samples. The training module is configured to train the initial tie-line load prediction model using multiple target training samples until the preset convergence condition is met, thus obtaining the trained tie-line load prediction model. The prediction module is configured to acquire the multi-dimensional data to be predicted for the date to be predicted and input it into the trained tie-line load prediction model, and output the tie-line load prediction result for the date to be predicted. The multi-dimensional data to be predicted includes boundary feature information and meteorological feature information for the date to be predicted.

[0008] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0010] Compared with existing technologies, the beneficial effects of this application's embodiments include at least the following: It overcomes the limitations of traditional tie-line load forecasting methods that rely on fixed power transmission and reception agreements, historical averages, and manual adjustments. By tracing back from the date to be predicted, it collects multiple historical multi-dimensional data points covering meteorological, temporal, and power supply-demand boundaries, and uses this historical multi-dimensional data to construct a multi-dimensional data-driven dynamic forecasting model for tie-line load. This model can systematically integrate the core dynamic influencing factors that affect and determine the size of tie-line power, replacing subjective experience-based judgments with a quantitative analysis mechanism, significantly improving the accuracy and reliability of the forecasting model, thereby meeting the actual needs of high-precision tie-line load forecasting in the construction of a unified inter-provincial power market. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic flowchart of a tie-line load forecasting method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a tie-line load prediction device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0014] A tie-line load prediction method and apparatus according to embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating a tie-line load forecasting method provided in an embodiment of this application. Figure 1 The tie-line load forecasting method can be executed by the server. For example... Figure 1 As shown, the tie-line load forecasting method includes: Step S101: Starting from the date to be predicted, collect multiple historical multi-dimensional data points backward; each historical multi-dimensional data point includes at least the historical boundary feature information and historical meteorological feature information of its corresponding historical date; Step S102: For each historical multi-dimensional data point, perform feature filtering and reconstruction on the historical multi-dimensional data to obtain the corresponding first candidate training sample; the total number of first candidate training samples is equal to the total number of historical multi-dimensional data points. Step S103: Select multiple target training samples from multiple first candidate training samples; wherein the total number of target training samples is less than or equal to the total number of first candidate training samples. Step S104: Use multiple target training samples to train the initial tie line load prediction model until the preset convergence condition is met, and obtain the trained tie line load prediction model. Step S105: Obtain the multi-dimensional data to be predicted corresponding to the date to be predicted and input it into the trained tie-line load prediction model, and output the tie-line load prediction result corresponding to the date to be predicted. The multi-dimensional data to be predicted includes boundary feature information and meteorological feature information corresponding to the date to be predicted.

[0016] Boundary feature information includes the unified dispatch load, interconnection line load, and renewable energy output of the province to be predicted and the provinces that interact with it (such as electricity), as well as the boundary features on both the supply and demand sides.

[0017] Meteorological characteristics information, including 100-meter wind speed, surface solar radiation downflow, temperature, humidity and other meteorological characteristics of the province to be measured and the provinces that interact with it by load (such as electricity).

[0018] Historical boundary feature information includes the boundary features on both the supply and demand sides of the province to be predicted and the provinces that interact with it (such as electricity) on the Mth day before the predicted date, such as the load of the unified dispatch, the load of the tie line, and the output of new energy; where M is a positive integer ≥1.

[0019] Historical meteorological characteristics information, including the 100-meter wind speed, surface solar radiation downflow, temperature, humidity and other meteorological characteristics of the province to be measured on the M-day before the forecast date and the provinces that interact with it by load (such as electricity).

[0020] The preset convergence condition can be that the number of iterations reaches a preset threshold (which can be set according to the actual situation, such as 50 rounds, 100 rounds, 200 rounds, etc.), or the target metric of the validation set (such as RMSE) does not decrease in K consecutive iterations.

[0021] In the training process of the tie-line load prediction model, the core purpose of setting convergence conditions is to balance model accuracy and training efficiency, and to avoid overfitting or ineffective iterations. In practical applications, appropriate convergence conditions can be set according to the model type and business requirements, and no specific restrictions are imposed in this application.

[0022] As an example, for the date to be predicted To mark the date, trace back to the first collection point. N Heaven (i.e.) N The number of data points (N can be set according to the actual situation, for example, it can be set to 120, 150, etc.) can be marked as historical multi-dimensional data. , , ... ,in, This indicates the day before the predicted date. This indicates the day before the predicted date. This indicates the 3rd day before the predicted date, and so on. This represents the Nth day before the date to be predicted. For example, the date to be predicted... October 10, 2025 October 9, 2025 October 8, 2025 October 7, 2025.

[0023] Historical multi-dimensional data Includes historical boundary features and historical meteorological features up to the day before the predicted date; historical multi-dimensional data. Includes historical boundary features and historical meteorological features from the day before the predicted date; historical multi-dimensional data. This includes historical boundary features and historical meteorological features from the 3rd day prior to the predicted date; and so on, historical multi-dimensional data. This includes historical boundary features and historical meteorological features from the Nth day prior to the predicted date.

[0024] The technical solution provided in this application overcomes the limitations of traditional tie-line load forecasting methods that rely on fixed power transmission and reception agreements, historical averages, and manual adjustments. By tracing back from the date to be predicted, it collects multiple historical multi-dimensional data points covering meteorological, temporal, and power supply-demand boundaries, and uses this data to construct a multi-dimensional data-driven dynamic forecasting model for tie-line load. This model systematically integrates the core dynamic influencing factors that determine the size of tie-line power, replacing subjective experience-based judgments with a quantitative analysis mechanism. This significantly improves the accuracy and reliability of the forecasting model, thereby meeting the practical needs of high-precision tie-line load forecasting in the construction of a unified inter-provincial power market.

[0025] In some embodiments, historical boundary feature information includes at least historical load information from the central dispatch system, historical tie-line load information, and renewable energy output information. Feature filtering and reconstruction of historical multi-dimensional data yields corresponding first candidate training samples, including: By filtering and reconstructing historical boundary feature information, combined features strongly correlated with tie line load are obtained; By filtering and reconstructing historical meteorological data, meteorological characteristics of the entire province that are strongly correlated with the load of the connecting lines were obtained. Based on historical tie-line load information, point-in-time lag instantaneous characteristics and periodic lag mean characteristics are constructed. The timestamp information of historical multi-dimensional data is quantized and encoded to obtain a time feature set; By combining the combined features, provincial meteorological features, time-lag instantaneous features, periodic lag mean features, and time feature sets, the first candidate training samples corresponding to historical multi-dimensional data are obtained.

[0026] The technical solution provided in this application, by performing feature filtering and reconstruction operations on each historical multi-dimensional data point, mines and forms a feature space that can profoundly reveal the physical nature and operational rules of tie line load. This feature space can accurately characterize the complex formation mechanism and monthly cycle characteristics of tie line load, enabling subsequent prediction models to learn feature information that is more strongly correlated with tie line load prediction, thus laying a data foundation for improving prediction accuracy.

[0027] In some embodiments, historical boundary feature information is filtered and reconstructed to obtain combined features strongly correlated with tie line load, including: Historical boundary characteristic information is divided into demand-side historical boundary characteristic information and supply-side historical boundary characteristic information; By performing a non-negative weighted linear combination of historical boundary characteristics of the demand side and the historical boundary characteristics of the supply side, a combined characteristic strongly correlated with tie line load is obtained.

[0028] Non-negative weighted linear combination refers to the operation of multiplying multiple variables (or features) by a weight coefficient greater than or equal to 0, then performing linear summation (or addition and subtraction) to obtain a new variable (i.e., combined feature).

[0029] As an example, firstly, historical boundary characteristic information is divided into demand-side historical boundary characteristic information and supply-side historical boundary characteristic information based on business attribute labels (such as demand or supply). Demand-side historical boundary characteristic information is marked as "positive" and directly retains the original value (or the original value after weighting) when participating in linear combination. Supply-side historical boundary characteristic information is marked as "negative" and is represented in the form of "subtraction" when participating in linear combination. Then, the 3σ principle or box plot method is used to filter the demand-side and supply-side historical boundary characteristic information respectively to remove outliers (such as a characteristic value exceeding the historical mean ± 3 times the standard deviation). After that, the filtered demand-side and supply-side historical boundary characteristic information is normalized to map all characteristic values ​​to the [0,1] interval to eliminate dimensional differences, resulting in the normalized values ​​of demand-side and supply-side historical boundary characteristics. Next, non-negative weights are assigned to each type of feature. The magnitude of the weight corresponds to the degree of influence of the feature on the tie-line load. The goal is to maximize the absolute value of the Pearson correlation coefficient between the combined new feature and the tie-line load, while ensuring that the weights conform to business logic (core features have higher weights, such as total regional load > new industrial load). For example, a linear regression method with non-negative constraints can be used, with the original feature values ​​of the same historical period as input and the actual tie-line load value as output, to train a linear regression model and solve for the non-negative optimal weights, ensuring that the correlation between the combined feature and the tie-line load is maximized. Next, the normalized values ​​of the historical boundary features on the demand side and the normalized values ​​of the historical boundary features on the supply side are linearly combined according to formula (1) to obtain the combined features.

[0030] (1); In equation (1), Indicates combined features; The weight value represents the normalized value of the i-th historical boundary feature of the demand side. This represents the normalized value of the i-th historical boundary feature of the demand side; The weight value represents the normalized value of the j-th supply-side historical boundary feature; This represents the normalized value of the j-th supply-side historical boundary feature.

[0031] The technical solution provided in this application directly models the fundamental driving force of inter-regional power flow by linearly combining historical boundary feature information on the demand side and the historical boundary feature information on the supply side, thus clearly quantifying the correlation between the power generation capacity at the sending end, the power demand at the receiving end, and the tie-line load. Furthermore, the use of a non-negative weighted linear combination method avoids the situation where negative weights violate business logic (e.g., negative demand feature weights would reverse the impact of supply and demand on the tie-line load).

[0032] In some embodiments, historical meteorological characteristic information is subjected to feature filtering and reconstruction to obtain provincial meteorological characteristics strongly correlated with the load of the connecting lines, including: Historical meteorological characteristic information is divided into meteorological characteristic information strongly correlated with new energy sources and general meteorological characteristic information; For meteorological feature information strongly correlated with new energy, the first reconstructed meteorological feature is obtained by weighting and reconstructing it according to the grid new energy installed capacity ratio of the province to be measured. For general meteorological feature information, area weighting is used for uniform reconstruction to obtain the second reconstructed meteorological features; The first and second reconstructed meteorological features were subjected to feature scale transformation and normalization to obtain the provincial meteorological features that are strongly correlated with the load of the connecting line.

[0033] Historical meteorological characteristic information, which is gridded coordinate meteorological data, refers to a set or more sets of meteorological element data with spatial coordinate attributes obtained by dividing the target geographical area (such as a province or the power grid coverage area of ​​a region) according to a preset latitude and longitude grid, and collecting or simulating them at each grid node at a specific time granularity (such as hour or 15 minutes).

[0034] As an example, a geographical region is divided into several uniformly sized rectangular grids (e.g., 0.1) based on latitude and longitude. ×0.1 The resolution is approximately 10km x 10km (corresponding to an actual geographical distance of about 10km x 10km). Each grid has a unique spatial coordinate identifier (such as the latitude and longitude of the grid center point), ensuring no omissions or overlaps in the coverage area. Each grid node corresponds to one or more sets of meteorological element data, including elements strongly correlated with new energy sources (i.e., meteorological characteristic information strongly correlated with new energy sources) and general meteorological elements (i.e., general meteorological characteristic information). Elements strongly correlated with new energy sources include 100m wind speed, surface solar radiation downflow, and wind direction, which directly affect wind and solar power output. General meteorological elements include temperature, humidity, air pressure, and precipitation, which indirectly affect power load and grid operation. All meteorological element data corresponds one-to-one with grid coordinates, accurately reflecting the spatial distribution differences of meteorological conditions within the region. Meteorological element data is collected or output at a fixed time granularity, forming a continuous time series, supporting the extraction of meteorological features at different time scales (daytime, medium- to long-term), maintaining consistency with the time granularity of tie-line load. For example, a provincial power grid divides its jurisdiction into 50 0.2 ×0.2 The latitude and longitude grid is used, and each grid node outputs a set of meteorological element data every 15 minutes. For example, the meteorological element data output by a certain grid node at a certain time point is: [118.5°E, 32.0°N, 100m wind speed = 6.8 m / s, surface solar radiation downflow = 610 W / m², temperature = 24.2℃, humidity = 55%].

[0035] For meteorological feature information strongly correlated with new energy, the first reconstructed meteorological feature is obtained by weighting and reconstructing according to the proportion of new energy installed capacity in the province to be tested. Specifically, firstly, the proportion of grid new energy installed capacity (such as grid wind power installed capacity and grid photovoltaic installed capacity) of the province to be tested can be calculated according to formula (2).

[0036] (2); In equation (2), This represents the proportion of new energy installed capacity in the k-th grid node, satisfying... ; This represents the total installed capacity of new energy sources (such as total installed wind power capacity or total installed photovoltaic capacity) of the k-th grid node. This represents the total number of grid nodes in the entire province; This represents the total installed capacity of new energy vehicles in the province.

[0037] Wind power is strongly correlated with 100m wind speed, and photovoltaic output is strongly correlated with the downward flux of solar radiation on the ground. Meteorological conditions in high-energy-generating areas have a greater impact on the province's energy output. Therefore, weighted reconstruction based on the grid's energy-generating capacity ratio can highlight the influence of core areas.

[0038] As an example, the meteorological characteristics of new energy strongly correlated with meteorological features (such as 100m wind speed and surface solar radiation downward flux) can be reconstructed by weighting the whole province according to formula (3) to obtain the first reconstructed meteorological features.

[0039] (3); In equation (3), This represents the equivalent wind speed (or equivalent solar radiation flux) of all grid nodes in the province at a certain point in time, which is the first reconstructed meteorological feature. This represents the 100 m wind speed value (or the downward flux of solar radiation on the ground surface) at a certain time point for the k-th grid node.

[0040] Factors such as temperature and humidity are not directly and strongly correlated with the output of new energy sources. They need to reflect the average meteorological conditions of the entire province. Therefore, they are reconstructed by weighting and uniformly reconstructing according to the grid area ratio.

[0041] As an example, the general meteorological feature information can be reconstructed by weighting the whole province according to formula (4) to obtain the second reconstructed meteorological features.

[0042] (4); In equation (4), This indicates the second reconstructed meteorological characteristic; This represents the percentage of the total area of ​​the province relative to the area of ​​the k-th grid node. satisfy ; It represents the values ​​of common meteorological elements (such as temperature, humidity, etc.).

[0043] Next, feature scaling transformations (such as linear, nonlinear, or logarithmic transformations) are performed on the first and second reconstructed meteorological features to improve the correlation between meteorological elements and tie-line loads. Specifically, the first and second reconstructed meteorological features are first nonnegated to avoid errors in logarithmic transformation. For example, if... If less than 0, take 0; if If the value is less than 0, then take the value 0.

[0044] Finally, the transformed first and second reconstructed meteorological features are normalized to eliminate the dimensional differences between the transformed first and second reconstructed meteorological features and the tie line load, enhance the correlation of the data, and obtain the provincial meteorological features that are strongly correlated with the tie line load.

[0045] The above scheme can transform gridded coordinate meteorological data into provincial meteorological characteristics that fit the regional new energy output characteristics, eliminate dimensional differences, and enhance the correlation between meteorological elements and the load of connecting lines.

[0046] By reconstructing the province's meteorological values ​​based on the grid-based renewable energy installed capacity ratio (such as the grid-based wind and solar installed capacity ratio), the original meteorological data (such as 100m wind speed, surface solar radiation downflow flux, etc.) are transformed into physical driving characteristics that can truly reflect the renewable energy power generation potential and thus affect the regional net load and power exchange demand, thereby improving the modeling accuracy of the impact of meteorological factors on tie line load.

[0047] In some embodiments, based on historical tie-line load information, point-in-time lag instantaneous characteristics and periodic lag mean characteristics are constructed, including: Extract multiple historical tie-line load values ​​corresponding to the target time point from historical tie-line load information to construct time-lag instantaneous features; Based on the preset periodic attributes, multiple historical tie-line load values ​​corresponding to the target time point are extracted from the historical tie-line load information to construct the periodic lag mean feature.

[0048] Historical tie-line load information, including timestamps (accurate to the hour or 15 minutes) and the historical tie-line load at the corresponding time point.

[0049] As an example, the specific steps for constructing point-in-time lag instantaneous features are as follows: First, calculate the offset step size corresponding to the lag days based on the time granularity of historical tie-line load information. For example, if the time granularity is hourly, then 1 day = 24 hours, and the offset step size for a lag of t days is t × 24; or, if the time granularity is 15 minutes, then 1 day = 96 time points, and the offset step size for a lag of t days is t × 96. Then, determine the range of lag days (e.g., the target time point). Extracted from historical tie-line load information (days 1-45 prior). The load values ​​of a single historical tie line at the same time point from day 1 to day 45 are used to retain single-point fluctuation information and generate time-lag instantaneous characteristics.

[0050] Periodic lagged mean characteristics include daily periodic lagged mean characteristics, weekly periodic lagged mean characteristics, and monthly periodic lagged mean characteristics.

[0051] The daily cyclical lag mean characteristic represents the historical tie-line load average at the same time point over the past L days, capturing the intraday tie-line load variation pattern. For example, the load average of 7 consecutive historical tie-line loads at the same hour over the past 7 days.

[0052] The weekly lag mean characteristic represents the historical tie-line load average at the same time on the same day of the past H weeks, capturing the variation pattern of tie-line load within a week. For example, the load average of four consecutive historical tie-line loads at the same time on the same Monday of the past four weeks.

[0053] The monthly lag mean characteristic represents the historical tie-line load average at the same time on similar dates in the past G months, capturing the changing patterns of monthly tie-line load. For example, the average load of three consecutive historical tie-line loads at the same time on the 1st ± 2 days of each of the past 3 months.

[0054] Preset periodic attributes, including daily, weekly, monthly, and yearly.

[0055] As an example, suppose we need to capture the historical tie line load change trend at the same time of day (such as the historical tie line load average at 12:00 every day). Then, based on the timestamp, we extract the historical tie line load values ​​corresponding to the same time point as the target time (12:00) (such as 12:00 every day for the past 7 days) from the historical tie line load information, and calculate the average value of these historical tie line load values ​​to obtain the daily periodic lag mean characteristic.

[0056] As an example, suppose we need to capture the load trend at the same time on the same workday within a week (such as the historical tie line load average at 12:00 every Monday). Then, based on the timestamp, we extract multiple historical tie line load values ​​corresponding to the same time point (12:00) in the historical tie line load information (such as 12:00 every Monday in the last 4 weeks), and calculate the average value of these historical tie line load values ​​to obtain the weekly lag mean characteristic.

[0057] As an example, suppose we need to capture the historical tie line load change trend at the same point in time (such as the historical tie line load average at 12:00 on the 1st of each month ± 2 days). Then, based on the timestamp, we extract multiple historical tie line load values ​​corresponding to the same period of the target time (12:00) in the historical tie line load information (such as the historical tie line load average at 12:00 on the 1st of each month ± 2 days in the past 3 months), and calculate the average value of these historical tie line load values ​​to obtain the monthly cycle lag mean characteristic.

[0058] In some embodiments, the time feature set includes time-dimensional encoded features, week-dimensional encoded features, month-dimensional encoded features, and year-dimensional encoded features. The timestamp information of historical multi-dimensional data is quantized and encoded to obtain the time feature set, which includes: The timestamp information is parsed to obtain time-dimension, week-dimension, month-dimension, and year-dimension information; Sine and cosine encoding is performed on the time-dimensional information to obtain the time-dimensional encoded features; One-hot encoding is performed on the weekly, monthly, and yearly information respectively to obtain the corresponding weekly, monthly, and yearly encoded features.

[0059] The time dimension information reflects the details of time within a day, and can be data at the 15-minute or hourly level. If it's 15-minute data, it includes not only the hour and minute but also the corresponding hour index value (e.g., 0~95, corresponding to 96 15-minute time points per day). If it's hourly data, it includes the hour (index value 0~23). Weekly dimension information, the attributes related to "week" correspond to "day of the week". Optional extraction of week number (to assist in analyzing weekly trends) aligns with the weekly cycle pattern of the connecting line load.

[0060] Monthly dimension information, attributes related to "month", corresponding to the month number. Optional extraction dates and number of days in the month, aligning with the monthly cycle patterns of the connection line load (seasons, trading plans, etc.).

[0061] Year-related information includes attributes associated with the year and corresponding year numbers. Optional extraction of year date sequences can aid in analyzing long-term trends.

[0062] As an example, the timestamp information is "2025-01-10 08:15:00". Parsing this timestamp information, the time dimension information is 8 hours and 15 minutes, and the corresponding time point index is 33, that is, the 33rd 15-minute point (08:15); the week dimension information is Friday, the second week of 2025; the month dimension information is January, the 10th day of the month; and the year dimension information is 2025, the 10th day of the year.

[0063] Because the time dimension (e.g., 96 15-minute time points in a day) is a periodic cyclical feature (returning to 0 after 95, and returning to 0 after 23), one-hot encoding loses the periodic correlation (e.g., the proximity between time point 95 and time point 0). Sine and cosine encoding, however, can map discrete time points to continuous two-dimensional vectors, preserving the cyclical periodicity and the sequential order of the time points. Therefore, for the time dimension information, sine and cosine encoding is used for quantization to obtain the time dimension encoded features. These time dimension encoded features include the time point index, the corresponding sine value, and the corresponding cosine value.

[0064] Week (0~6), month (1~12), and year (e.g., 2025~2026) are discrete categorical features with no continuous numerical correlation (e.g., month 1 and 2 are categorical differences, not numerical differences). One-hot encoding can map each category to a binary column (1 indicates belonging to the category, 0 indicates not belonging), avoiding the model from misinterpreting the category order as numerical order, while retaining all category information.

[0065] For tie-line load forecasting, the business logic changes monthly, so the weight of the one-hot encoded features for each month needs to be increased to highlight their monthly variation patterns. The principle for adjusting feature weights is to keep the weights of other features at 1 (the baseline weight), and increase the weight of the one-hot encoded features for each month to a reasonable value greater than 1 (e.g., 2 or 3, which needs to fit the business scenario and avoid overfitting the model to the monthly features). One way to adjust feature weights is to directly multiply and scale the values ​​of the one-hot encoded columns for each month (i.e., multiply by a weight coefficient). This is simple, efficient, and results in no information loss, meeting the requirements for engineering implementation. A weight coefficient of 2-5 is recommended, with fixed weights preferred. Dynamic weights can be used in complex scenarios. The core is to enhance the impact of monthly features on the model and align with the monthly variation patterns of the tie-line business.

[0066] The above scheme, by giving special weight to the "monthly" feature, enables the model to explicitly learn and follow the inherent operational pattern of monthly adjustments to the tie line plan, accurately capture the impact of monthly plan step changes and seasonal trends on the load, and improve the accuracy of tie line load forecasting.

[0067] The aforementioned combined features, provincial meteorological features, point-lagged instantaneous features, period-lagged mean features, and time feature set together constitute a feature space that can profoundly reveal the physical nature and operational rules of the tie line load. This enables models (such as the XGBoost model) to learn operational rules far more deeply than traditional methods, significantly improving prediction accuracy.

[0068] In some embodiments, selecting multiple target training samples from multiple first candidate training samples includes: Outlier training samples are removed from multiple first candidate training samples to obtain multiple second candidate training samples, wherein the total number of second candidate training samples is less than or equal to the total number of first candidate training samples. Calculate the similarity between each pair of second candidate training samples, and select multiple target training samples from multiple second candidate training samples based on the similarity, wherein the total number of target training samples is less than or equal to the total number of second candidate training samples.

[0069] Considering that tie-line load data may contain outliers due to unplanned adjustments, blockages, or faults, this application introduces a dual data cleaning mechanism before model training: First, outlier data cleaning based on the 3σ principle (three standard deviations principle) identifies and removes abnormal samples that deviate from the normal range of data distribution; second, in-sample cosine similarity verification selects valid samples (i.e., target training samples) whose feature similarity conforms to the operating logic of the power system. These mechanisms effectively filter out noise and abnormal event interference, achieving purification of the training set.

[0070] As an example, outlier training samples are eliminated from multiple first-candidate training samples based on the 3σ principle. or ,in, Let represent any candidate training sample. This represents the mean of all candidate training samples. (where represents the standard deviation of all candidate training samples) to obtain multiple second candidate training samples. Next, the cosine similarity between each pair of second candidate training samples is calculated. If the mean of the cosine similarity between a second candidate training sample and other second candidate training samples is lower than a preset threshold, then that second candidate training sample is removed.

[0071] By optimizing the selection process for target training samples, the overall quality of training samples is improved; at the same time, the robustness of the model training process is enhanced to ensure that the prediction model still has stable performance and generalization ability under complex scenarios such as fluctuations in electricity market supply and demand and sudden weather changes.

[0072] In some embodiments, multiple target training samples obtained in the above embodiments are used to train an initial tie-line load prediction model (such as an XGBoost model). The core parameters of the XGBoost algorithm are set as follows: learning rate 0.05~0.3, maximum tree depth 3~8, number of decision trees 100~600, regularization coefficient L1=0.1~0.5, feature sampling rate 0.5~1, and sample sampling rate 0.5~1. The above parameter range is set as the parameter network. The multiple target training samples are divided into training set and validation set in a ratio of 8:2. The RMSE (root mean square error) of the prediction results of the validation set is used as the loss function. The parameters are iterated 50 times based on the Bayesian optimization algorithm to determine the optimal parameter combination. The model is trained according to the optimal parameters until the preset convergence condition is reached, and the trained tie-line load prediction model is obtained.

[0073] By employing Bayesian optimization algorithms, automated hyperparameter optimization is performed within a parameter space specifically designed to improve model generalization ability (such as controlling the depth of the decision tree and introducing L1 / L2 regularization coefficients).

[0074] The above-described model training process ensures that the final model not only fits well to common historical operating patterns, but also has excellent inference (generalization) capabilities for complex operating scenarios that have not been experienced but are reasonable, thus guaranteeing the stability and reliability of the prediction results.

[0075] For the multi-dimensional data to be predicted corresponding to the date to be predicted, the same feature engineering as in the above embodiment is used to form an input vector, and the input vector is input into the above-trained tie-line load prediction model for inference, and the tie-line load prediction result corresponding to the date to be predicted is output.

[0076] After the basic prediction model has been trained and rolling prediction samples have been accumulated, the following error correction prediction process will be executed: First, based on the accumulated training sample set, the true value of the connectline load for each sample is extracted, along with the predicted value of the connectline load output by the basic prediction model. The difference between the two is calculated as the prediction residual. Then, using the full feature set after feature engineering (i.e., multiple target training samples) as input features, and the calculated prediction residual as the model output target, an error correction model based on gradient boosting trees is constructed. The model training parameters are fixed as follows: tree depth max_depth=3, learning rate learning_rate=0.05, number of decision trees n_estimators=400, with other parameters using default configurations. The error correction model is trained based on the training sample set. Finally, the feature engineering results of the samples to be predicted (i.e., the multi-dimensional data to be predicted) are input into the trained error correction model to obtain the residual prediction value of the samples to be predicted. The predicted value of the samples to be predicted output by the basic prediction model is added to the residual prediction value to obtain the final connectline load prediction result.

[0077] In addition to the core prediction model, this application innovatively introduces an independent error correction model. This model uses the historical prediction residuals of the core prediction model as the learning target, and by mining the temporal distribution patterns and feature correlation patterns of the residuals, it automatically identifies and corrects the systematic biases or inherent errors that may exist in the core model under different operating scenarios.

[0078] The core prediction model and the error correction model work together to form a closed-loop prediction system with self-improvement capabilities. This system not only provides high-precision initial prediction results, but also continuously iterates and optimizes model parameters and correction strategies by accumulating historical operating data, achieving continuous evolution of prediction performance, maintaining long-term technological leadership, and providing stable and reliable technical support for power market participants.

[0079] In summary, the beneficial effects of the technical solutions provided in this application include at least the following: (1) By systematically integrating the power grid operation data of both sending and receiving provinces, and by constructing a linear combination that reflects the supply and demand balance, the coupling characteristics that directly drive inter-regional power exchange are established, thereby fundamentally modeling the formation mechanism of tie-line loads. The power grid operation data of both sides includes power generation output data of sending end, power load data of receiving end, capacity limit data of inter-regional transmission channels, and power market transaction data; the linear combination of supply and demand balance is constructed based on the quantitative calculation of the remaining power generation capacity of sending end and the power gap of receiving end, directly mapping the core driving source of tie-line loads, enabling the model to accurately capture the essential laws of inter-regional power flow.

[0080] (2) Based on the meteorological data utility reconstruction technology of the ratio of wind and solar new energy installed capacity at the sending and receiving ends, meteorological conditions are transformed into physical quantities that affect the net load and exchange demand of the region. At the same time, the time-series coding of monthly cycle weights is strengthened, and the model influence of monthly characteristics is enhanced so that the prediction results strictly follow the scheduling and operation rules of the tie line plan adjusted monthly.

[0081] To address the deficiency that raw meteorological data (wind speed, light intensity, temperature, precipitation, etc.) cannot directly reflect the potential of new energy power generation, this application reconstructs meteorological data by weighting it according to the proportion of wind power and photovoltaic installed capacity at the sending and receiving ends, generating meteorological characteristics for the entire province. These meteorological characteristics are used as physical quantities to directly correlate with the net load of the region (total load - new energy power generation), thereby accurately depicting the impact of meteorological conditions on the demand for interconnection line exchange.

[0082] By increasing the weight of monthly features in model training, the model can explicitly learn the scheduling rules for monthly adjustments to the connection plan. Specifically, monthly features are one-hot encoded during the feature preprocessing stage and given a higher weight coefficient than regular time-series features, ensuring that the prediction results match the actual operational changes in the monthly plan.

[0083] (3) Before model training, the original training set is cleaned and purified by a combination of outlier sample screening based on the 3σ principle and internal consistency verification of the sample set based on cosine similarity, so as to construct high-quality training samples and improve the stability and generalization starting point of model training.

[0084] Calculate the mean and standard deviation of each feature dimension in the original training data, and remove feature values ​​exceeding [μ]. Outlier samples in the range of [3σ, μ+3σ] can be effectively filtered to remove abnormal data interference caused by unplanned adjustments, power grid congestion, or equipment failure.

[0085] Calculate the cosine similarity between each pair of second candidate samples, select target training samples whose similarity meets the power system operation logic threshold, and eliminate abnormal samples whose feature correlation violates physical laws to ensure the consistency and rationality of data distribution within the training set.

[0086] The training samples purified by the above-mentioned dual verification method eliminate the interference of noisy data on the model, lay a high-quality data foundation for subsequent model training, and significantly improve the initial generalization ability of the model.

[0087] (4) After the mainstream prediction model outputs the results, an error correction model with historical prediction residuals as labels and the same feature engineering as input is trained to perform secondary calibration on the initial prediction results, forming an end-to-end prediction system that can be self-iteratively optimized, thereby achieving a systematic improvement in prediction accuracy.

[0088] The combined features processed by feature engineering, the meteorological features of the whole province, the instantaneous features with time lag, the mean features with periodic lag, and the time feature set are input into the trained tie-line load prediction model (such as the XGBoost model) to obtain the tie-line load prediction results.

[0089] Collect the historical prediction residuals (actual value - initial prediction value) of the trained tie-line load prediction model (core prediction model) as labels, use the feature vectors output by the same feature engineering as input, and train the error correction model; use this model to correct the deviation of the initial prediction value output by the core prediction model to obtain the final prediction result.

[0090] As prediction data continues to accumulate, new prediction residuals and corresponding feature data are incorporated into the training set of the error correction model, enabling iterative updates of model parameters. This allows the entire prediction system to have self-optimization capabilities and maintain a high level of prediction accuracy over the long term.

[0091] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0092] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0093] Figure 2 This is a schematic diagram of a tie-line load prediction device provided in an embodiment of this application. Figure 2 As shown, the tie-line load forecasting device 200 includes: The data acquisition module 201 is configured to collect multiple historical multi-dimensional data points backward from the date to be predicted; each historical multi-dimensional data point includes at least the historical boundary feature information and historical meteorological feature information of its corresponding historical date. The first screening module 202 is configured to perform feature screening and reconstruction on each piece of historical multi-dimensional data to obtain the corresponding first candidate training sample; the total number of first candidate training samples is equal to the total number of historical multi-dimensional data. The second filtering module 203 is configured to filter out multiple target training samples from multiple first candidate training samples; wherein the total number of target training samples is less than or equal to the total number of first candidate training samples. Training module 204 is configured to train the initial tie-line load prediction model using multiple target training samples until the preset convergence condition is met, thereby obtaining the trained tie-line load prediction model. The prediction module 205 is configured to acquire the multi-dimensional data to be predicted corresponding to the date to be predicted and input it into the trained tie-line load prediction model, and output the tie-line load prediction result corresponding to the date to be predicted. The multi-dimensional data to be predicted includes boundary feature information and meteorological feature information corresponding to the date to be predicted.

[0094] In some embodiments, historical boundary characteristic information includes at least historical load information of the central dispatch system, historical tie-line load information, and renewable energy output information.

[0095] The first screening module 202 mentioned above includes: The first screening and reconstruction unit is configured to perform feature screening and reconstruction on historical boundary feature information to obtain combined features that are strongly correlated with tie line load. The second screening and reconstruction unit is configured to perform feature screening and reconstruction on historical meteorological feature information to obtain the provincial meteorological features that are strongly correlated with the load of the connecting line. The building unit is configured to construct point-in-time lag instantaneous features and periodic lag mean features based on historical tie-line load information; The encoding unit is configured to quantize and encode the timestamp information of historical multi-dimensional data to obtain a time feature set. The combination unit is configured to combine the combined features, provincial meteorological features, point-lagged instantaneous features, periodically lagged mean features, and time feature sets to obtain the first candidate training sample corresponding to historical multi-dimensional data.

[0096] In some embodiments, the first filtering and reconstruction unit described above may be specifically configured as follows: Historical boundary characteristic information is divided into demand-side historical boundary characteristic information and supply-side historical boundary characteristic information; By performing a non-negative weighted linear combination of historical boundary characteristics of the demand side and the historical boundary characteristics of the supply side, a combined characteristic strongly correlated with tie line load is obtained.

[0097] In some embodiments, the second filtering and reconstruction unit described above may be specifically configured as follows: Historical meteorological characteristic information is divided into meteorological characteristic information strongly correlated with new energy sources and general meteorological characteristic information; For meteorological feature information strongly correlated with new energy, the first reconstructed meteorological feature is obtained by weighting and reconstructing it according to the grid new energy installed capacity ratio of the province to be measured. For general meteorological feature information, area weighting is used for uniform reconstruction to obtain the second reconstructed meteorological features; The first and second reconstructed meteorological features were subjected to feature scale transformation and normalization to obtain the provincial meteorological features that are strongly correlated with the load of the connecting line.

[0098] In some embodiments, the above-mentioned building unit may be specifically configured as follows: Extract multiple historical tie-line load values ​​corresponding to the target time point from historical tie-line load information to construct time-lag instantaneous features; Based on the preset periodic attributes, multiple historical tie-line load values ​​corresponding to the target time point are extracted from the historical tie-line load information to construct the periodic lag mean feature.

[0099] In some embodiments, the time feature set includes time-dimensional encoding features, week-dimensional encoding features, month-dimensional encoding features, and year-dimensional encoding features. The aforementioned encoding unit can be specifically configured as follows: The timestamp information is parsed to obtain time-dimension, week-dimension, month-dimension, and year-dimension information; Sine and cosine encoding is performed on the time-dimensional information to obtain the time-dimensional encoded features; One-hot encoding is performed on the weekly, monthly, and yearly information respectively to obtain the corresponding weekly, monthly, and yearly encoded features.

[0100] In some embodiments, the second filtering module 203 described above may be specifically configured as follows: Outlier training samples are removed from multiple first candidate training samples to obtain multiple second candidate training samples, wherein the total number of second candidate training samples is less than or equal to the total number of first candidate training samples. Calculate the similarity between each pair of second candidate training samples, and select multiple target training samples from multiple second candidate training samples based on the similarity, wherein the total number of target training samples is less than or equal to the total number of second candidate training samples.

[0101] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0102] Figure 3 This is a schematic diagram of the electronic device 300 provided in an embodiment of this application. For example... Figure 3 As shown, the electronic device 300 of this embodiment includes a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the various device embodiments described above.

[0103] Electronic device 300 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 300 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 300 and does not constitute a limitation on electronic device 300. It may include more or fewer parts than shown, or different parts.

[0104] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0105] The memory 302 can be an internal storage unit of the electronic device 300, such as a hard disk or RAM of the electronic device 300. The memory 302 can also be an external storage device of the electronic device 300, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 300. The memory 302 can also include both internal and external storage units of the electronic device 300. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0106] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments 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 unit can be implemented in hardware or as a software functional unit.

[0107] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium can be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0108] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A tie-line load forecasting method, characterized in that, include: Starting from the date to be predicted, multiple historical multi-dimensional data points are collected backwards; each of these historical multi-dimensional data points includes at least historical boundary feature information and historical meteorological feature information for its corresponding historical date. For each piece of historical multi-dimensional data, feature filtering and reconstruction are performed on the historical multi-dimensional data to obtain the corresponding first candidate training sample; the total number of the first candidate training samples is equal to the total number of the historical multi-dimensional data. Multiple target training samples are selected from multiple first candidate training samples; wherein the total number of target training samples is less than or equal to the total number of first candidate training samples; The initial tie-line load prediction model is trained using multiple target training samples until the preset convergence condition is met, thus obtaining the trained tie-line load prediction model. The multi-dimensional data to be predicted for the date to be predicted is obtained and input into the trained tie-line load prediction model, and the tie-line load prediction result for the date to be predicted is output. The multi-dimensional data to be predicted includes boundary feature information and meteorological feature information corresponding to the date to be predicted.

2. The method according to claim 1, characterized in that, The historical boundary feature information includes at least historical load information of the central dispatch system, historical load information of the tie line, and new energy output information; The historical multi-dimensional data is subjected to feature filtering and reconstruction to obtain the corresponding first candidate training samples, including: The historical boundary feature information is filtered and reconstructed to obtain combined features that are strongly correlated with tie line load; The historical meteorological characteristics information is filtered and reconstructed to obtain the provincial meteorological characteristics that are strongly correlated with the load of the connecting lines. Based on the historical tie-line load information, point-in-time lag instantaneous characteristics and periodic lag mean characteristics are constructed. The timestamp information of the historical multi-dimensional data is quantized and encoded to obtain a time feature set; The combined features, provincial meteorological features, time-lag instantaneous features, periodic lag mean features, and time feature set are combined to obtain the first candidate training sample corresponding to the historical multidimensional data.

3. The method according to claim 2, characterized in that, The historical boundary feature information is filtered and reconstructed to obtain combined features strongly correlated with tie line load, including: The historical boundary feature information is divided into demand-side historical boundary feature information and supply-side historical boundary feature information; The historical boundary characteristics of the demand side and the historical boundary characteristics of the supply side are combined linearly with non-negative weights to obtain combined characteristics that are strongly correlated with tie line load.

4. The method according to claim 2, characterized in that, The historical meteorological characteristics information is filtered and reconstructed to obtain provincial meteorological characteristics strongly correlated with the load of the connecting lines, including: The historical meteorological feature information is divided into meteorological feature information strongly correlated with new energy sources and general meteorological feature information; The meteorological feature information strongly correlated with new energy is reconstructed by weighting according to the grid new energy installed capacity ratio of the province to be tested, and the first reconstructed meteorological feature is obtained. The general meteorological feature information is uniformly reconstructed using area weighting to obtain the second reconstructed meteorological feature; The first and second reconstructed meteorological features were subjected to feature scale transformation and normalization to obtain the provincial meteorological features that are strongly correlated with the load of the connecting line.

5. The method according to claim 2, characterized in that, Based on the historical tie-line load information, point-in-time lag instantaneous characteristics and periodic lag mean characteristics are constructed, including: Extract multiple historical tie-line load values ​​corresponding to the target time point from the historical tie-line load information to construct time-lag instantaneous features; According to the preset periodic attributes, multiple historical tie-line load values ​​corresponding to the target time point are extracted from the historical tie-line load information to construct the periodic lag mean feature.

6. The method according to claim 2, characterized in that, The time feature set includes time-dimensional encoding features, week-dimensional encoding features, month-dimensional encoding features, and year-dimensional encoding features; The timestamp information of the historical multi-dimensional data is quantized and encoded to obtain a time feature set, including: The timestamp information is parsed to obtain time-dimension information, week-dimension information, month-dimension information, and year-dimension information; The time-dimensional information is subjected to sine and cosine encoding to obtain the time-dimensional encoded features; One-hot encoding is performed on the weekly, monthly, and yearly dimension information respectively to obtain the corresponding weekly, monthly, and yearly dimension encoding features.

7. The method according to claim 1, characterized in that, Multiple target training samples are selected from the multiple first candidate training samples, including: Outlier training samples are removed from multiple first candidate training samples to obtain multiple second candidate training samples, wherein the total number of second candidate training samples is less than or equal to the total number of first candidate training samples. Calculate the similarity between each pair of second candidate training samples, and select multiple target training samples from multiple second candidate training samples based on the similarity, wherein the total number of target training samples is less than or equal to the total number of second candidate training samples.

8. A tie-line load prediction device, characterized in that, include: The data acquisition module is configured to collect multiple historical multi-dimensional data points backward from the date to be predicted; each of the historical multi-dimensional data points includes at least historical boundary feature information and historical meteorological feature information for its corresponding historical date. The first filtering module is configured to perform feature filtering and reconstruction on each piece of historical multi-dimensional data to obtain a corresponding first candidate training sample; the total number of the first candidate training samples is equal to the total number of the historical multi-dimensional data. The second filtering module is configured to filter out multiple target training samples from multiple first candidate training samples; wherein the total number of target training samples is less than or equal to the total number of first candidate training samples; The training module is configured to train the initial tie-line load prediction model using multiple target training samples until a preset convergence condition is met, thereby obtaining a trained tie-line load prediction model. The prediction module is configured to acquire the multi-dimensional data to be predicted for the date to be predicted and input it into the trained tie-line load prediction model, and output the tie-line load prediction result for the date to be predicted. The multi-dimensional data to be predicted includes boundary feature information and meteorological feature information corresponding to the date to be predicted.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images