Semantic-enhanced data-knowledge fusion method and system for short-term load forecasting of a transformer area

Abstract

A semantic enhanced data-knowledge fusion substation short-term load prediction method and system, collects the multi-source prediction data of the target substation and carries out structured data preprocessing, obtains structured prediction characteristics and converts into numerical semantic vectors, constructs semantic time series characteristics, converts semantic vectors into semantic time series characteristics matched with the prediction task, obtains structured prediction characteristics and effective semantic characteristics after screening and fusion, forms data-knowledge fusion prediction characteristics and trains a prediction model, inputs the data-knowledge fusion prediction characteristics corresponding to the to-be-predicted day into the trained load prediction model in the online stage, obtains the short-term load prediction result of the target substation and carries out evaluation. The present application converts the semantic disturbance information in the external event text into calculable, screenable and fusible prediction characteristics and models with traditional prediction variables such as historical load, weather and calendar, thereby improving the accuracy, stability and complex scene adaptability of substation short-term load prediction.

Description

Technical Field

[0001] This invention relates to a technology in the field of power system control, specifically a semantically enhanced data-knowledge fusion method and system for short-term load forecasting of distribution areas. Background Technology

[0002] Existing methods for short-term load forecasting in transformer substations largely rely on structured data such as historical load data, meteorological data, and calendar information, making it difficult to effectively utilize unstructured textual information such as news, policies, announcements, and public events. When external events cause abnormal load fluctuations, existing technologies struggle to address the challenges of quantifying external event text, aligning it with time, and the impact of invalid semantic variables on forecast stability. Summary of the Invention

[0003] This invention addresses the aforementioned shortcomings of existing technologies by proposing a semantically enhanced data-knowledge fusion method and system for short-term load forecasting of power distribution areas. It transforms semantic perturbation information from external event texts into computable, filterable, and fusionable predictive features, which are then modeled together with traditional predictive variables such as historical load, weather, and calendar data. This improves the accuracy, stability, and adaptability to complex scenarios in short-term load forecasting of power distribution areas. It solves the problem that traditional methods for short-term load forecasting of power distribution areas primarily rely on structured data such as historical load, weather, and calendar data, making it difficult to perceive the impact of external factors such as policies, news, and public events on load changes. Through stationarity testing and causal screening mechanisms, the reliability and interpretability of semantic features are improved. Furthermore, through collaborative modeling of structured data and unstructured knowledge information, the accuracy, stability, and adaptability to complex scenarios in short-term load forecasting of power distribution areas are enhanced.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a semantically enhanced data-knowledge fusion method for short-term load forecasting of a target transformer area. The method involves collecting multi-source forecasting data of the target transformer area and performing structured data preprocessing to obtain structured forecasting features. These features are then converted into numerical semantic vectors, and semantic time-series features are constructed. These semantic vectors are further converted into semantic time-series features that match the forecasting task. After screening, structured forecasting features and effective semantic features are obtained and fused to form data-knowledge fusion forecasting features. A forecasting model is then trained. During the online phase, the data-knowledge fusion forecasting features corresponding to the forecast date are input into the trained load forecasting model to obtain and evaluate the short-term load forecasting results for the target transformer area.

[0006] This invention relates to a short-term load forecasting system for a transformer substation area implementing the above-mentioned method, comprising: a multi-source data acquisition module, a structured data preprocessing module, a text semantic encoding module, a semantic time series construction module, a semantic feature filtering module, a data-knowledge fusion module, a load forecasting module, and a forecasting evaluation module, sequentially connected to form a complete data processing and load forecasting link; wherein: the multi-source data acquisition module acquires historical load data, meteorological data, calendar information, and event text data of the target transformer substation area; the structured data preprocessing module preprocesses the historical load data, meteorological data, and calendar information to obtain structured forecasting features; the text semantic encoding module performs text cleaning, word segmentation, term filtering, and... The semantic vector of the text is obtained through word frequency-inverse document frequency semantic encoding. The semantic time series construction module performs daily-scale aggregation and lag feature construction on the text semantic vector to obtain semantic time series features. The semantic feature filtering module performs ADF stationarity test and bidirectional Granger causality filtering on the semantic time series features to obtain effective semantic features. The data-knowledge fusion module fuses the structured prediction features with the effective semantic features to obtain data-knowledge fusion prediction features. The load prediction module inputs the data-knowledge fusion prediction features into the preset load prediction model and outputs the short-term load prediction results for the target transformer area. The prediction evaluation module evaluates the short-term load prediction results and outputs the prediction error analysis results.

[0007] The technical effects of this invention include:

[0008] 1. This invention establishes a mechanism for event text time assignment and sample alignment for short-term load forecasting of distribution areas. Addressing the issues of irregular release times of news, policies, announcements, and public events, and the inconsistency between event occurrence times and text release times, this invention assigns event texts to corresponding forecast dates based on their release time, occurrence time, or validity period, and establishes a correspondence with daily-scale forecast samples for the target distribution area. This allows previously discrete and irregular event texts to be incorporated into the data construction process for next day's load curve forecasting.

[0009] 2. This invention constructs a method for generating event semantic time series based on "text semantic vector—daily-scale semantic intensity—lag semantic features". Specifically, the event text is first cleaned, segmented, filtered, and semantically encoded using term frequency-inverse document frequency to form a text semantic vector; then, the text semantic vectors within the same date are aggregated at the daily scale to obtain daily-scale semantic intensity features; finally, lag semantic features are constructed by combining a preset lag window. This method can transform unstructured event text into semantic time series features that are consistent with the time scale of load prediction samples and can characterize the impact of event delays.

[0010] 3. This invention proposes a two-stage screening mechanism for semantic time series data for transformer area load forecasting. Specifically, the ADF stationarity test is first used to eliminate non-stationary semantic features that do not meet the requirements of time series analysis. Then, a two-way Granger causality test is used to determine the predictive lead relationship between the semantic features and the target transformer area load sequence, retaining only semantic features that have effective predictive information on load changes. This mechanism can remove weakly correlated, insignificant, or spuriously correlated text variables before semantic features enter the prediction model, reducing the interference of invalid semantic information on load forecasting results.

[0011] 4. This invention establishes a data-knowledge fusion prediction sample construction method for fine-grained load curve prediction of the next day. Specifically, preprocessed historical load features, meteorological features, and calendar features are combined with effective semantic features after daily-scale aggregation, lag construction, and statistical screening. Sample index matching and feature concatenation are performed according to a unified prediction date to form data-knowledge fusion prediction features. The load curves of the target area at multiple time sections of the next day are used as prediction targets. This method achieves a unified expression of structured numerical data and external event semantic information in the same prediction sample.

[0012] 5. This invention constructs an integrated method for semantic feature generation, fusion modeling, and forecast output for short-term load forecasting of power distribution areas. Specifically, the method sequentially performs semantic encoding, daily-scale aggregation, lag feature construction, and effective semantic feature selection on event text to form semantic time-series features capable of characterizing the impact of external events. These semantic time-series features are then uniformly concatenated with historical load features, meteorological features, and calendar features to construct a data-knowledge fusion forecast sample. Finally, the fusion forecast sample is used to train the load forecasting model, and the load curve for the next day is output based on the fusion forecast features corresponding to the date to be predicted. This method realizes a complete technical chain from semantic extraction and feature selection of event text information to load forecasting applications.

[0013] 6. This invention establishes a grouped prediction and error comparison mechanism for verifying the contribution of effective semantic features. Specifically, under the same data partitioning and prediction model conditions, load-only input model, load-calendar input model, load-calendar-meteorological input model, and load-calendar-meteorological-semantic input model are constructed respectively. The prediction results of each model are evaluated by mean absolute error, root mean square error, and symmetric mean absolute percentage error. This mechanism can analyze the prediction differences and error changes before and after the introduction of semantic features, thereby clarifying the contribution of effective semantic features to the short-term load prediction accuracy, curve tracking ability, and daily-scale error stability of the distribution area. Attached Figure Description

[0014] Figure 1 This is a flowchart of the present invention;

[0015] Figure 2 This is a schematic diagram of the system structure for an example embodiment;

[0016] Figure 3 This is a schematic diagram comparing the actual load curve with the predicted load curves of different feature combinations in an example.

[0017] Figure 4 This is a schematic diagram comparing the daily-scale mean absolute error of the structured baseline model and the semantic enhancement model in an example. Detailed Implementation

[0018] like Figure 1 As shown in this embodiment, a semantically enhanced data-knowledge fusion method for short-term load forecasting of distribution areas includes:

[0019] S100, acquire multi-source prediction data for the target transformer area;

[0020] The target distribution areas include: residential distribution areas, industrial and commercial distribution areas, industrial park distribution areas, distribution transformer power supply areas, or other distribution units with independent load metering data.

[0021] The multi-source prediction data includes two categories: structured data and unstructured data. Structured data includes historical load data, meteorological data, and calendar information. Unstructured data includes event text data related to the area where the target station is located.

[0022] The historical load data includes: active load sequence, reactive load sequence, load peak, load trough, daily load curve, load change rate, historical load on the same day, and historical load on adjacent days for the target transformer area within a historical period.

[0023] The meteorological data includes: temperature, humidity, wind speed, rainfall, air pressure, weather type, perceived temperature, maximum temperature, minimum temperature, and average temperature.

[0024] The calendar information includes: date, hour, minute, weekday type, weekday identifier, weekend identifier, holiday identifier, workday adjustment identifier, season information, and month information.

[0025] The event text data includes: regional news texts, policy announcement texts, public event texts, social activity texts, industrial operation texts, traffic incident texts, emergency texts, meteorological disaster early warning texts, and park announcement texts.

[0026] The historical load data is obtained from the power distribution automation system, the electricity consumption information collection system, or the intelligent terminal of the transformer area; the meteorological data is obtained from the meteorological service platform or the regional meteorological monitoring system; the calendar information is automatically generated by the system according to the date rules; and the event text data is obtained through the web page collection interface, the government information disclosure platform, the news platform, the announcement system, or the manually maintained knowledge base.

[0027] S200 performs structured data preprocessing on historical load data, meteorological data, and calendar information to obtain structured forecast features, specifically including:

[0028] S201 identifies missing values ​​in historical load data and meteorological data: if the number of consecutive missing points is less than a preset threshold, linear interpolation, nearby time imputation, historical similar day imputation, or mean imputation methods are used for repair; if the number of consecutive missing points is greater than a preset threshold, the corresponding sample is marked as a low-quality sample or removed from the training sample.

[0029] S202, Identification and correction of outliers: After identifying outliers using methods such as box plots, standard deviation, quantile thresholds, or detection methods based on the rate of change between adjacent time points, the identified abnormal load points or abnormal weather points are corrected by interpolation between adjacent time points, replacement with historical similar days, or replacement with model estimates.

[0030] S203, Time alignment of data from different sources: The forecast task is to forecast the load curve at a 5-minute resolution for the next day, so there are 288 forecast points per day.

[0031] S204 involves normalizing or standardizing continuous variables to reduce the impact of dimensional differences on the training of the predictive model. The normalization method used is min-max normalization, and the standardization method uses mean-variance standardization.

[0032] S205 reconstructs the preprocessed historical load, meteorological, and calendar information into daily forecast samples, so that each sample corresponds to a next-day load curve forecasting task.

[0033] S300 cleans and semantically encodes the event text data, converting the original text into a numerical semantic vector, specifically including:

[0034] S301, perform text cleaning on the original text, including: removing web page tags, invalid symbols, duplicate text, advertising content, special characters, and stop words.

[0035] S302, perform word segmentation on the cleaned text, and correct the word segmentation results based on a preset dictionary, domain thesaurus, or a custom power domain keyword library.

[0036] S303 filters the segmented terms, removing terms that appear too infrequently, have weak semantic contribution, or are obviously irrelevant, thus forming a candidate term set.

[0037] S304, calculates the weight of each term in the text based on the term frequency-inverse document frequency method. For the nth term in the mth document, its TF-IDF weight is... ,in: For the first The first in the text Term frequency-inverse document frequency weighting for each term; For the first The term in the first Word frequency in the text; For including the first The number of text terms; This represents the total number of texts in the text collection.

[0038] S305, according to each term Weights are used to construct text semantic vectors. For any given... The semantic vector of a text ,in: For the first The semantic vector of a text; This represents the number of terms retained after term filtering.

[0039] S400, constructing semantic time series features, transforms text semantic vectors into semantic time series features that match the prediction task, specifically including:

[0040] S401, assign the event text to the corresponding date based on the publication time, occurrence time, or effective time of the event text.

[0041] S402 aggregates one or more text semantic vectors within the same date to obtain the daily-scale semantic intensity features for that date.

[0042] The aggregation methods employed include summation, mean, weighted mean, maximum value, or weighted aggregation based on text importance. For example, let... Given the set of texts corresponding to date t, then the semantic strength features of date t are... ,in: Represents the daily-scale semantic intensity features of date t; This represents the semantic vector of the m-th text within date t; Aggregate represents the aggregation function.

[0043] S402, considering the potential lag effect of external events on load changes, construct lag semantic features. For a preset maximum lag order L, constructing semantic lag features transforms irregularly published event texts into semantic time series consistent with the time scale of load forecast samples, enabling textual semantic information to participate in short-term load forecasting for transformer areas. Specifically: ,in: This represents the semantic time series features corresponding to date t; This represents the semantic intensity feature with a lag of 1 day.

[0044] S500 performs statistical filtering on semantic time series features, specifically including:

[0045] S501, ADF stationarity test: Perform an ADF stationarity test on each semantic time series feature. If a semantic feature sequence fails the stationarity test, it is considered that the semantic feature has strong non-stationarity, which may lead to unreliable statistical relationships, and it is removed from the candidate feature set.

[0046] S502, Two-way Granger Causality Screening: Semantic features that pass the ADF stationarity test are subjected to a two-way Granger causality test with the target area load sequence. This test is used to determine whether historical semantic information can provide additional predictive information for future load changes.

[0047] If a semantic feature has a statistically significant predictive lead relationship with the load sequence, then the semantic feature is retained; if the semantic feature does not show a significant predictive lead relationship, or only shows a reverse effect of the load on the semantic feature, then the semantic feature is removed.

[0048] By using the ADF stationarity test and bidirectional Granger causality screening, more reliable and predictive effective semantic features can be selected from high-dimensional text semantic features, thereby improving the robustness and interpretability of subsequent data-knowledge fusion prediction models.

[0049] S600, after obtaining structured prediction features and effective semantic features, fuses the two to form data-knowledge fusion prediction features and trains the prediction model, specifically including:

[0050] S601, by horizontally concatenating historical load features, meteorological features, calendar features, and effective semantic features, a complete prediction input vector is obtained: ,in: For the data-knowledge fusion prediction features corresponding to date t; Historical load characteristics; It is a calendar feature; Meteorological characteristics; These are valid semantic features. The corresponding prediction target is the next day's load curve: ,in: This represents the load curve for the next day; H represents the number of forecast points within a day. H=288.

[0051] S602 divides historical samples into training, validation, and test sets. The training set is used to train the load prediction model, multiple random decision trees are constructed, and the output results are integrated and averaged. The validation set is used to adjust the model parameters, and the test set is used to evaluate the model's prediction performance.

[0052] The load prediction model mentioned above uses extreme random tree model, random forest model, gradient boosting tree model, support vector regression model, long short-term memory network model, temporal convolutional network model or Transformer prediction model.

[0053] S700 outputs and evaluates short-term load forecast results. Specifically, it inputs the data-knowledge fusion forecast features corresponding to the forecast date into the trained load forecast model, obtains the short-term load forecast results for the target transformer area, and then evaluates the forecast results using mean absolute error, root mean square error, and symmetric mean absolute percentage error.

[0054] The short-term load forecast result is the load curve of 288 points at a 5-minute resolution for the target distribution area the following day.

[0055] The mean absolute error (MAE) Root Mean Square Error (RMSE) Symmetric Mean Absolute Percentage Error (SMAPE) ,in: This represents the actual load value at the i-th time point; H represents the predicted load value at the i-th time point; H represents the number of prediction points.

[0056] Preferably, to analyze the contribution of different data sources to the prediction effect, four types of comparative models are constructed, including: the first type of model is a load-only input model, which uses only historical load data; the second type of model is a load-calendar input model, which uses historical load data and calendar information; the third type of model is a load-calendar-meteorological input model, which uses historical load data, calendar information and meteorological data; and the fourth type of model is a load-calendar-meteorological-semantic input model, which uses historical load data, calendar information, meteorological data and effective semantic features.

[0057] The gain of semantic features on the short-term load forecast results of the transformer area is quantified by comparing the prediction errors of the above models.

[0058] like Figure 2As shown, this embodiment relates to a data-knowledge fusion system for short-term load forecasting of distribution areas that implements the above-mentioned method. The system includes: a multi-source data acquisition module, a structured data preprocessing module, a text semantic encoding module, a semantic time series construction module, a semantic feature filtering module, a data-knowledge fusion module, a load forecasting module, and a forecasting and evaluation module. The multi-source data acquisition module is connected to both the structured data preprocessing module and the text semantic encoding module. Both the structured data preprocessing module and the text semantic encoding module are connected to the semantic time series construction module. The semantic time series construction module, the semantic feature filtering module, the data-knowledge fusion module, the load forecasting module, and the forecasting and evaluation module are connected sequentially, ultimately outputting the next day's load curve and error evaluation results.

[0059] The structured data preprocessing module includes an alignment unit, a standardization unit, and an organization unit. The alignment unit performs time alignment for data with different time resolutions according to a unified prediction granularity. The standardization unit uses max-min normalization or mean-variance standardization methods to eliminate dimensional differences for continuous variables. The organization unit organizes the processed historical load, meteorological, and calendar data into daily samples, so that each sample corresponds to a next-day load curve prediction task.

[0060] The text semantic encoding module includes a cleaning unit, a word segmentation correction unit, a filtering unit, and a vector construction unit. Specifically: the cleaning unit performs deduplication, noise reduction, stop word removal, and invalid character cleanup on the event text; the word segmentation correction unit performs word segmentation on the cleaned text and corrects the segmentation results using a power industry thesaurus, a regional event thesaurus, or a user-defined thesaurus; the filtering unit filters terms, retaining candidate terms that contribute semantically; and the vector construction unit calculates the semantic weight of each term based on the TF-IDF method and constructs a text semantic vector.

[0061] The semantic time series construction module includes a time mapping unit, an aggregation unit, and a feature construction unit. The time mapping unit maps the text semantic vector to the corresponding date based on the event text's publication time, occurrence time, or effective time. The aggregation unit uses aggregation methods such as summation, mean, weighted mean, or maximum value for multiple event texts within the same date to form daily-scale semantic intensity features. The feature construction unit constructs one or more semantic features with different lag orders based on a preset lag window to characterize the delayed impact of external events on the load changes in the transformer area, obtaining semantic time series features consistent with the time scale of the load prediction sample.

[0062] The semantic feature screening module includes a stationarity test unit and a causality test unit. The stationarity test unit performs an ADF stationarity test on each dimension of semantic time series features and removes semantic features that do not meet the stationarity requirements. The causality test unit performs a bidirectional Granger causality test on the semantic features that pass the stationarity test and the target area load sequence to determine whether the historical semantic information can provide predictive lead information for future load changes. Semantic features with significant predictive lead relationships are retained, while semantic features with weak correlations, insignificant correlations, or only pseudo-correlation relationships are removed.

[0063] The load forecasting module employs extreme random tree models, random forest models, gradient boosting tree models, support vector regression models, long short-term memory network models, temporal convolutional network models, or Transformer prediction models to model the data-knowledge fusion prediction features, outputting a fine-grained load curve for the target transformer area for the next day. With a prediction time resolution of 5 minutes, this module outputs load forecast values ​​for 288 time segments for the next day.

[0064] The prediction and evaluation module includes a deviation calculation unit and a comparison analysis unit. The deviation calculation unit calculates the deviation between the predicted load curve and the actual load curve based on the mean absolute error, root mean square error, and symmetric mean absolute percentage error indicators. The comparison analysis unit compares and analyzes the prediction models under different feature combinations to quantify the contribution of semantic features to the prediction accuracy and analyzes the tracking effect of the representative daily load curve and the trend of daily-scale error changes to evaluate the stability of the model under complex disturbance scenarios.

[0065] The comparative analysis includes: load-only input model, load-calendar input model, load-calendar-meteorological input model, and load-calendar-meteorological-semantic input model.

[0066] Through practical application experiments, the semantically enhanced data-knowledge fusion method and system for short-term load forecasting of power distribution areas, based on this invention, was run offline in a computer environment with data processing and machine learning model training capabilities. The experimental object was a next-day fine-grained load curve forecasting task for a target power distribution area, with a forecast time resolution of 5 minutes and 288 forecast points per day. The experimental data included historical load data, meteorological data, calendar information, and text data of regional news, policy announcements, and public events for the target power distribution area. The test period covered continuous forecast samples from January 2025 to June 2025. In the experiment, firstly, missing value processing, outlier correction, normalization, and time alignment were performed on historical load, meteorological, and calendar information to obtain structured prediction features. Secondly, text cleaning, word segmentation, term filtering, and TF-IDF semantic encoding were performed on the event text data, and daily-scale aggregation and lag feature construction were carried out according to the release time, occurrence time, or effective time of the event text. Then, the constructed semantic time series features were subjected to ADF stationarity test and bidirectional Granger causality screening to obtain effective semantic features. Finally, under the same training set, validation set, and test set partitioning method, load-only input model, load-calendar input model, load-calendar-meteorological input model, and load-calendar-meteorological-semantic input model were constructed respectively, and the prediction effect was evaluated by mean absolute error, root mean square error, and symmetric mean absolute percentage error. Experimental results show that the model using only historical load features has a MAE of 184.50, an RMSE of 256.83, and a SMAPE of 3.98%. After adding calendar information, the MAE decreases to 177.37, the RMSE to 245.64, and the SMAPE to 3.83%. Further addition of meteorological information reduces the MAE to 161.76, the RMSE to 220.93, and the SMAPE to 3.54%. Finally, adding effective semantic features further reduces the MAE to 93.59, the RMSE to 125.27, and the SMAPE to 2.05%. Therefore, the effective semantic features obtained through event text semantic encoding, diurnal-scale aggregation, lag construction, and semantic feature selection in this invention can further reduce the prediction error of the next-day load curve for the substation area. The experimental data are shown in Table 1 below.

[0067] To verify the impact of different feature combinations on the prediction results, four sets of comparative models were set up in this specific application example: Model D: using only historical load features; Model D+C: using historical load features and calendar features; Model D+C+W: using historical load features, calendar features, and meteorological features; Model D+C+W+S: using historical load features, calendar features, meteorological features, and effective semantic features. All four models adopted the same prediction model structure and the same training and testing partitioning method, ensuring that the differences in prediction performance mainly stemmed from the different combinations of input features. The prediction results were evaluated using mean absolute error, root mean square error, and symmetric mean absolute percentage error. The prediction results of different models are shown in the table below.

[0068] Table 1 Comparison of overall prediction accuracy of prediction models with different feature combinations

[0069] As shown in Table 1, the model prediction error gradually decreases as calendar information, meteorological information, and semantic features are gradually incorporated. The D model, which only uses historical load data, has a relatively large prediction error. After adding calendar information, the D+C model can further characterize the load differences under different date types such as weekdays, weekends, and holidays. After adding meteorological information, the D+C+W model can further reflect the impact of environmental factors such as temperature, humidity, and weather type on load changes. Based on this, after further incorporating effective semantic features, the D+C+W+S model achieves the best prediction results, with a MAE of 93.59, an RMSE of 125.27, and a SMAPE of 2.05%.

[0070] Compared to the strongest structured baseline model D+C+W, the semantically enhanced model D+C+W+S of this invention reduced its MAE from 161.76 to 93.59 and its RMSE from 220.93 to 125.27, representing reductions of 42.14% and 43.30%, respectively. This indicates that the event text features, after semantic construction and statistical filtering, can provide effective predictive information beyond traditional structured variables, thereby significantly improving the accuracy of short-term load forecasting for transformer substations.

[0071] Furthermore, in the analysis of the representative daily forecast curve, the semantically enhanced model can more accurately track the changing trend of the actual load curve, especially during periods of rapid load increase, local turning points, and large fluctuations, where its forecast curve is closer to the actual curve. Compared to structured baseline models that only use historical load, calendar, and meteorological variables, the semantically enhanced model has a stronger ability to characterize load disturbances driven by external events.

[0072] Furthermore, in the daily-scale error trend analysis, the semantic enhancement model maintained a low daily average error throughout the testing period, indicating that the prediction improvement brought by semantic features was not limited to individual samples but exhibited good stability under different date conditions. Therefore, this invention not only improves overall prediction accuracy but also enhances the ability to track representative daily load curves and the stability of daily-scale prediction errors.

[0073] Compared to existing methods for short-term load forecasting of substation areas that only utilize structured features, the performance improvement of this device / method lies in the following: Through processes such as event text time attribution, TF-IDF semantic encoding, diurnal-scale aggregation, lag feature construction, ADF stationarity testing, and bidirectional Granger causality screening, regional news, policy announcements, and public event texts are transformed into effective semantic features that match the substation load forecasting samples. These effective semantic features are then integrated with historical load, calendar information, and meteorological data to form data-knowledge fusion forecasting features. Based on this technical link, the load-calendar-meteorological-semantic input model, compared to a structured baseline model that only uses historical load, calendar information, and meteorological data, reduces MAE from 161.76 to 93.59, RMSE from 220.93 to 125.27, and SMAPE from 3.54% to 2.05%; specifically, MAE is reduced by 42.14%, and RMSE by 43.30%. The above results show that the newly added event text semantic time series construction and effective semantic feature screening steps of this invention can reduce the interference of invalid text variables on the prediction model, and improve the prediction accuracy and error stability of the next day's fine-grained load curve of the transformer area under complex disturbance scenarios.

[0074] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A semantically enhanced data-knowledge fusion method for short-term load forecasting of distribution areas, characterized in that, Multi-source forecast data of the target transformer area is collected and structured data preprocessing is performed to obtain structured forecast features. After converting them into numerical semantic vectors, semantic time series features are constructed. The semantic vectors are then converted into semantic time series features that match the forecast task. After screening, structured forecast features and effective semantic features are obtained and fused to form data-knowledge fusion forecast features. The forecast model is then trained. During the online phase, the data-knowledge fusion forecast features corresponding to the forecast day are input into the trained load forecast model to obtain the short-term load forecast results of the target transformer area and evaluate them. The target distribution areas include: residential distribution areas, industrial and commercial distribution areas, industrial park distribution areas, distribution transformer power supply areas, or other distribution units with independent load metering data; The multi-source forecast data includes two categories: structured data and unstructured data. Structured data includes historical load data, meteorological data, and calendar information; unstructured data includes event text data related to the area where the target station is located. The historical load data includes: the active load sequence, reactive load sequence, peak load, valley load, daily load curve, load change rate, historical load on the same day, and historical load on adjacent days for the target transformer area within a historical period; The meteorological data includes: temperature, humidity, wind speed, rainfall, air pressure, weather type, perceived temperature, maximum temperature, minimum temperature, and average temperature. The calendar information includes: date, hour, minute, weekday type, weekday identifier, weekend identifier, holiday identifier, workday adjustment identifier, season information, and month information; The event text data includes: regional news texts, policy announcement texts, public event texts, social activity texts, industry operation texts, traffic incident texts, emergency incident texts, meteorological disaster early warning texts, and park announcement texts; The historical load data is obtained from the power distribution automation system, the electricity consumption information collection system, or the intelligent terminal of the transformer area; the meteorological data is obtained from the meteorological service platform or the regional meteorological monitoring system; the calendar information is automatically generated by the system according to the date rules; and the event text data is obtained through the web page collection interface, the government information disclosure platform, the news platform, the announcement system, or the manually maintained knowledge base.

2. The semantically enhanced data-knowledge fusion method and system for short-term load forecasting of distribution areas according to claim 1, characterized in that, specifically... include: S100, acquire multi-source prediction data for the target transformer area; S200 performs structured data preprocessing on historical load data, meteorological data, and calendar information to obtain structured forecast features, specifically including: S300 cleans and semantically encodes the event text data, converting the original text into a numerical semantic vector, specifically including: S400, constructing semantic time series features, transforms text semantic vectors into semantic time series features that match the prediction task, specifically including: S401, assign the event text to the corresponding date based on the publication time, occurrence time, or effective time of the event text; S402, aggregate one or more text semantic vectors within the same date to obtain the daily-scale semantic intensity features of that date; S402, considering the potential lag effect of external events on load changes, a lag semantic feature is constructed. For a preset maximum lag order L, the semantic lag feature transforms irregularly published event texts into semantic time series consistent with the time scale of load forecast samples, enabling text semantic information to participate in short-term load forecasting for transformer areas. Specifically: ,in: This represents the semantic time series features corresponding to date t; The semantic intensity feature representing a lag of 1 day; S500 performs statistical filtering on semantic time series features. S600, after obtaining structured prediction features and effective semantic features, fuses the two to form data-knowledge fusion prediction features and trains the prediction model, specifically including: S601, by horizontally concatenating historical load features, meteorological features, calendar features, and effective semantic features, a complete prediction input vector is obtained: ,in: For the data-knowledge fusion prediction features corresponding to date t; Historical load characteristics; It is a calendar feature; Meteorological characteristics; For effective semantic features, the corresponding prediction target is the next day's load curve: ,in: This is the load curve for the next day; H is the number of forecast points within a day. S602 divides historical samples into training set, validation set and test set, uses training set to train load prediction model, builds multiple random decision trees and integrates and averages the output results, uses validation set to adjust model parameters, and uses test set to evaluate model prediction effect. S700 outputs and evaluates short-term load forecast results. Specifically, it inputs the data-knowledge fusion forecast features corresponding to the forecast date into the trained load forecast model, obtains the short-term load forecast results for the target transformer area, and then evaluates the forecast results using mean absolute error, root mean square error, and symmetric mean absolute percentage error.

3. The semantically enhanced data-knowledge fusion method and system for short-term load forecasting of distribution areas according to claim 2, characterized in that, The S200 specifically includes: S201, Identify missing values ​​in historical load data and meteorological data: If the number of consecutive missing points is less than a preset threshold, linear interpolation, nearby time imputation, historical similar day imputation, or mean imputation methods are used for repair; if the number of consecutive missing points is greater than a preset threshold, the corresponding sample is marked as a low-quality sample or removed from the training sample. S202, Identification and correction of outliers: After identifying outliers using methods such as box plots, standard deviation, quantile thresholds, or detection methods based on the rate of change between adjacent time points, the identified abnormal load points or abnormal weather points are corrected by interpolation between adjacent time points, replacement with historical similar days, or replacement with model estimates. S203, Time alignment of data from different sources: The forecasting task is to forecast the load curve at a 5-minute resolution for the next day, so there are 288 forecast points per day; S204. Normalize or standardize continuous variables to reduce the impact of dimensional differences on the training of the prediction model. The normalization method is max-min normalization, and the standardization method is mean-variance standardization. S205 reconstructs the preprocessed historical load, meteorological, and calendar information into daily forecast samples, so that each sample corresponds to a next-day load curve forecasting task.

4. The semantically enhanced data-knowledge fusion method and system for short-term load forecasting of distribution areas according to claim 2, characterized in that, The S300 specifically includes: S301, perform text cleaning on the original text, including: removing web page tags, invalid symbols, duplicate text, advertising content, special characters and stop words; S302, perform word segmentation on the cleaned text, and correct the word segmentation results based on a preset dictionary, domain thesaurus, or a custom power domain keyword library; S303, filter the segmented terms and remove terms that appear too infrequently, have weak semantic contribution, or are obviously irrelevant, to form a candidate term set; S304, calculate the weight of each term in the text based on the term frequency-inverse document frequency method. For the nth term in the mth text, its TF-IDF weight is... ,in: For the first The first in the text Term frequency-inverse document frequency weighting for each term; For the first The term in the first Word frequency in the text; For including the first The number of text terms; This represents the total number of texts in the text collection. S305, according to each term Weights are used to construct text semantic vectors, for any _th_... The semantic vector of a text ,in: For the first The semantic vector of a text; This represents the number of terms retained after term filtering.

5. The semantically enhanced data-knowledge fusion method and system for short-term load forecasting of distribution areas according to claim 2, characterized in that, The S500 specifically includes: S501, ADF stationarity test: Perform an ADF stationarity test on each semantic time series feature. If a semantic feature sequence fails the stationarity test, it is considered that the semantic feature has strong non-stationarity, which may lead to unreliable statistical relationships, and it is removed from the candidate feature set. S502, Two-way Granger Causality Screening: For semantic features that pass the ADF stationarity test, a two-way Granger causality test is performed with the target area load sequence. This test is used to determine whether historical semantic information can provide additional predictive information for future load changes.

6. A data-knowledge fusion system for short-term load forecasting of distribution areas, implementing the method of any one of claims 1-5, characterized in that, include: The system comprises a multi-source data acquisition module, a structured data preprocessing module, a text semantic encoding module, a semantic time series construction module, a semantic feature filtering module, a data-knowledge fusion module, a load forecasting module, and a forecasting and evaluation module. Specifically, the multi-source data acquisition module is connected to both the structured data preprocessing module and the text semantic encoding module. Both the structured data preprocessing module and the text semantic encoding module are connected to the semantic time series construction module. The semantic time series construction module, the semantic feature filtering module, the data-knowledge fusion module, the load forecasting module, and the forecasting and evaluation module are connected sequentially, ultimately outputting the next day's load curve and error evaluation results.

7. The data-knowledge fusion distribution area short-term load forecasting system according to claim 6, characterized in that, The semantic time series construction module includes a time mapping unit, an aggregation unit, and a feature construction unit. The time mapping unit maps the text semantic vector to the corresponding date based on the event text's publication time, occurrence time, or effective time. The aggregation unit uses aggregation methods such as summation, mean, weighted mean, or maximum value for multiple event texts within the same date to form daily-scale semantic intensity features. The feature construction unit constructs one or more semantic features with different lag orders based on a preset lag window to characterize the delayed impact of external events on the load changes in the transformer area, obtaining semantic time series features consistent with the time scale of the load prediction sample.

8. The data-knowledge fusion distribution area short-term load forecasting system according to claim 6, characterized in that, The semantic feature screening module includes a stationarity test unit and a causality test unit. The stationarity test unit performs an ADF stationarity test on each dimension of semantic time series features and removes semantic features that do not meet the stationarity requirements. The causality test unit performs a bidirectional Granger causality test on the semantic features that pass the stationarity test and the target area load sequence to determine whether the historical semantic information can provide predictive lead information for future load changes. Semantic features with significant predictive lead relationships are retained, while semantic features with weak correlations, insignificant correlations, or only pseudo-correlation relationships are removed.

9. The data-knowledge fusion distribution area short-term load forecasting system according to claim 6, characterized in that, The load forecasting module uses extreme random tree model, random forest model, gradient boosting tree model, support vector regression model, long short-term memory network model, temporal convolutional network model or Transformer forecasting model to model the data-knowledge fusion forecasting features and outputs the fine-grained load curve of the target area for the next day. With a forecasting time resolution of 5 minutes, the module outputs the load forecast values ​​of 288 time sections for the next day.

10. The data-knowledge fusion distribution area short-term load forecasting system according to claim 6, characterized in that, The prediction and evaluation module includes a deviation calculation unit and a comparison analysis unit. The deviation calculation unit calculates the deviation between the predicted load curve and the actual load curve based on the mean absolute error, root mean square error, and symmetrical mean absolute percentage error indicators. The comparison analysis unit compares and analyzes the prediction models under different feature combinations to quantify the contribution of semantic features to the prediction accuracy and analyzes the tracking effect of the representative daily load curve and the trend of daily-scale error changes to evaluate the stability of the model under complex disturbance scenarios. The comparative analysis includes: load-only input model, load-calendar input model, load-calendar-meteorological input model, and load-calendar-meteorological-semantic input model.

Images