Deep learning based short-term power load forecasting system

Abstract

The short-term power load prediction system based on deep learning comprises the following steps: classifying, setting a sliding window and detecting integrity of local multi-source data uploaded by each edge node; detecting historical synchronous deviation of complete data, cross-validating cross-source data or combining real-time trend verification results for linkage determination; eliminating complete data marked as inaccurate data; performing two-dimensional coding processing and time distance factor embedding on the complete data to generate enhanced date feature combinations; performing multi-scale date feature interaction analysis based on the enhanced date feature combinations to output a scenario-based time reference; extracting cross-lag effect features of numerical time series corresponding to load data and driving factors in the complete data to obtain driving lag feature sequences; and obtaining load prediction values in the next load prediction time scale based on the scenario-based time reference and the driving lag feature sequences. The application accurately captures the load law and driving relationship, and significantly improves the prediction accuracy and robustness.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically a short-term power load forecasting system based on deep learning. Background Technology

[0002] As power systems evolve towards higher proportions of renewable energy grid integration and higher penetration rates of power electronic equipment, the randomness and volatility of electricity user loads have significantly increased. Accurate load forecasting has become a core supporting technology for ensuring the safe and stable operation of the power grid and improving energy efficiency. Currently, short-term power load forecasting systems based on deep learning have been widely researched and applied, but the following prominent technical shortcomings still exist: 1. At the data level, it is difficult to achieve secure data sharing and precise quality control across entities. There is a contradiction between data silos and privacy leaks. Furthermore, due to the lack of differentiated integrity detection and blockchain traceability mechanisms, data defects are missed during critical periods and abnormal data is difficult to trace. Low-quality data directly affects the accuracy of predictions. 2. In terms of feature representation, traditional date encoding cannot handle time cycle characteristics and multi-scale combination effects, and it treats historical data indiscriminately, making it difficult for the model to capture scene-specific load patterns such as "Friday evenings in summer" and to accurately focus on recent high-value time series information. 3. In terms of modeling and prediction, it ignores the cross-lag effect of power load and external factors, makes it difficult to quantify the delayed impact of driving factors such as temperature and electricity price, lacks a dynamic benchmark update mechanism, and cannot adapt to long-term trend deviations such as the expansion of user electricity consumption. A single model is difficult to take into account scenario-based, driving, and trend-based load fluctuations, which makes predictions prone to lag or bias and difficult to support diversified engineering application needs such as grid hierarchical dispatch and demand response. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide a short-term power load forecasting system based on deep learning, comprising the following steps: Step s1: Perform classification sliding window settings and integrity checks on the local multi-source data uploaded by each edge node, and determine whether the local multi-source data is complete based on the integrity check results; Step s2: Perform historical same-period deviation detection on the complete data, determine whether there is any suspicious load based on the detection results, and if so, perform cross-source data cross-validation on the suspicious load. Determine whether to combine the cross-validation results with the real-time trend verification results for linkage judgment based on the cross-validation results, and remove the complete data marked as inaccurate data. Step s3: Perform two-dimensional encoding and time distance factor embedding on the complete data to generate enhanced date feature combinations. Based on the enhanced date feature combinations, perform multi-scale date feature interaction analysis and output a scenario-based time benchmark. Step s4: Extract cross-lag effect features from the numerical time series of load data and driving factors in the complete data to obtain the driving lag feature sequence. Based on the scenario-based time base and the driving lag feature sequence, obtain the load forecast value in the next load forecast time scale.

[0004] Furthermore, a cloud data center is constructed, and edge nodes are deployed in each data subject within a preset scope. The cloud data center communicates with each edge node, and the edge nodes are used to de-identify the collected local multi-source data before uploading it to the cloud data center. The cloud data center is equipped with several blockchain nodes, which are interconnected to form a blockchain network. Local multi-source data uploaded by edge nodes is linked to the blockchain nodes through a data uploading terminal. The data uploading terminal is used to store the local multi-source data uploaded to the cloud data center to the blockchain nodes and mark the uploading time.

[0005] Furthermore, the process of classifying and performing sliding window settings and integrity checks on the local multi-source data uploaded by each edge node includes: The collection cycle, load prediction time scale, and weight level of different data types in the local multi-source data of each edge node are preset, the basic window length of different data types is set, and the basic window length of different data types is calibrated in a scenario-based manner based on the data body of each edge node. Based on the basic window length of the scenario-based calibration, sliding windows of different data types are set for each edge node. The window missing rate of sliding windows of different data types is obtained. Based on the window missing rate of sliding windows of different data types and the weight level of different data types, the integrity score of sliding windows is obtained. A preset integrity score threshold is set. If the integrity score of the sliding window is less than the integrity score threshold, the local multi-source data in the sliding window is marked as missing data. If the integrity score of the sliding window is greater than or equal to the integrity score threshold, the local multi-source data in the sliding window is marked as complete data.

[0006] Furthermore, the process of detecting historical contemporaneous deviations in the complete data includes: Extract complete data from the historical synchronous window of the current sliding window of the edge node from the blockchain node, mark the complete data in the historical synchronous window as historical synchronous data, compare the complete data of the current sliding window with the complete data of the historical synchronous data, and obtain the load deviation rate of the load data and the synchronous deviation rate of other data types in the complete data. A preset load deviation rate threshold is set. Load data with a load deviation rate greater than the load deviation rate threshold is marked as suspicious loads. Cross-source data cross-validation is performed on suspicious loads. If the suspicious load is verified to have reasonable driving factors, the suspicious load is removed from the label. If the suspicious load is found to have no reasonable driving factors, the suspicious load is marked and then combined with the real-time trend verification results for a joint judgment.

[0007] Furthermore, the process of cross-validating suspicious loads across data sources includes: Obtain the data type that has potential driving logic with the load data, and mark the data type as the driving type; Collect complete data from each blockchain node within several historical sliding windows. Based on the driving type, extract driving features and load deviation rate from the complete data within each historical sliding window to obtain load deviation rate and driving features. Construct a feature regression model and use the load deviation rate and driving features as training data to train the feature regression model to obtain the trained feature regression model. The feature regression model that has been trained outputs the proportion of each driving type that explains the load deviation rate. The thresholds corresponding to the same period deviation rate and the explanation proportion are preset. If the same period deviation rate and the explanation proportion of the driving type are both greater than the corresponding thresholds, the driving type is marked as a driving factor. A reasonable deviation mapping table is constructed, which includes the load deviation rate range corresponding to the same period deviation rate range of different driving factors. The total reasonable deviation rate range is obtained based on the same period deviation rate, explanation ratio and reasonable deviation mapping table of each driving factor. It is then determined whether the load deviation rate of the suspicious load is within the total reasonable deviation rate range. If it is, the suspicious load is determined to have reasonable driving factors. If it is not, the suspicious load is determined to have no reasonable driving factors.

[0008] Furthermore, the process of making a linkage judgment based on real-time trend verification results includes: Extract the load change rate of the current sliding window of the edge node and the historical maximum load change rate. Compare the load change rate with the historical maximum load change rate. If the load change rate is greater than the historical maximum load change rate, mark the complete data as inaccurate data and remove the inaccurate data. If the load change rate is less than or equal to the historical maximum load change rate, then extract multiple sets of historical data from the current sliding window and determine whether there is a deviation trend in the multiple sets of historical data from the same period. If a deviation trend exists, obtain the historical benchmark data of the same period for the current sliding window based on the deviation trend, and then compare the complete data of the current sliding window with the complete data of the historical benchmark data. If there is no bias trend, then mark the complete data as inaccurate data and remove the inaccurate data.

[0009] Furthermore, the process of performing two-dimensional encoding on the complete data and embedding time distance factors to generate enhanced date feature combinations includes: Extract the load data and driving factors corresponding to the numerical time series and date data from the complete data of each blockchain node within several historical sliding windows. Perform two-dimensional encoding on the date data within each historical sliding window to convert it into multiple two-dimensional embedding vectors. Merge the two-dimensional embedding vectors into a joint date vector. By setting a lag period, extracting the numerical time series corresponding to the load data within the historical sliding window and lag period where the joint date vector is located, and labeling the numerical time series corresponding to the load data within the historical sliding window and lag period as the core label of the joint date vector, embedding the time distance factor into the joint date vector after completing the core label labeling, and generating an enhanced date feature combination.

[0010] Furthermore, the process of performing multi-scale date feature interaction analysis based on enhanced date feature combinations to output a contextualized time reference includes: The enhanced date feature combination and the core labels of each joint date vector in the enhanced date feature combination are used as training data. A date scene prediction model is constructed based on the time-delay attention mechanism. The date scene prediction model is trained using the training data. The load value corresponding to different date feature combinations is output according to the trained date scene prediction model. The load values ​​corresponding to different combinations of date features are marked as scenario-based time bases.

[0011] Furthermore, the process of obtaining load forecast values ​​for the next load forecast timescale includes: Obtain the load data and the corresponding numerical time series of driving factors from the complete data of the current sliding window of the target blockchain node, and extract the cross-lag effect features from the load data and the corresponding numerical time series of driving factors to obtain the driving lag feature sequence. A user load prediction model is constructed based on the RBF neural network. Complete data from several historical collection periods stored in the target blockchain node is used as training data to train the user load prediction model. The user load prediction model is trained by inputting the complete data of the current sliding window of the target blockchain node, the driving lag feature sequence, and the scenario-based time reference. Based on the user load prediction model, the load prediction value of the target blockchain node in the next load prediction time scale is output.

[0012] Furthermore, the process of extracting cross-lag effect features from the numerical time series corresponding to load data and driving factors includes: Construct a cross-lag information analysis module and a lag attention module, and extract the load data and the numerical time series corresponding to the driving factors from the complete data of several historical sliding windows of each blockchain node as training data; Based on the lag period, the load data and the numerical time series corresponding to the driving factors are aligned by lag time series. The aligned load data and the numerical time series corresponding to the driving factors are then input into the cross-lag information analysis module. The lag embedding matrix is ​​output from the cross-lag information analysis module. The training data and the lag embedding matrix are input into the lag attention module. The lag attention module performs a linear transformation on the training data to generate the corresponding Q vector. Then, it maps the lag embedding matrix and the training data into different K and V vectors. The scaling dot product attention mechanism is used to obtain the weights between the Q vector and different K vectors. The V vector is weighted and summed to output the driving lag feature sequence.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1. By setting a categorized sliding window and dividing the data into weighted levels, differentiated integrity assessments are implemented for the core layer (such as peak load data and real-time temperature data), important layer, and basic layer data. Combined with scenario-based calibration (reducing the window length for critical periods by 50%), data missing issues are accurately captured, avoiding the omission of data defects in critical periods due to traditional "one-size-fits-all" detection. At the same time, the integrity score calculation based on weights highlights the key value of core data for prediction while also taking into account the basic supporting role of non-core data, ensuring that the data input to the model has high integrity and high relevance.

[0014] 2. Blockchain enables data security and traceability, meeting compliance requirements. Cloud data centers build blockchain networks to upload de-identified data from edge nodes to the chain in real time and mark it with a timestamp, realizing the immutability and traceability of data throughout its entire lifecycle from collection, transmission to storage.

[0015] 3. By employing a dual mechanism of historical period deviation detection and cross-source data cross-validation, combined with the XGBoost feature regression model to quantify the explanatory power of driving factors on load deviation, this method accurately distinguishes between "reasonable load fluctuations" and "inaccurate data." Compared to traditional methods that rely solely on a single threshold to determine anomalies, this approach avoids misclassifying load deviations caused by reasonable factors such as increased user electricity consumption or extreme weather as anomalies. It also effectively identifies and eliminates inaccurate data caused by data acquisition errors or equipment malfunctions, ensuring the purity of the dataset input to the prediction model and reducing the interference of anomalous data on model training and inference.

[0016] 4. Two-dimensional encoding overcomes the shortcomings of traditional date encoding, restoring the essential periodicity of time. Addressing the limitations of traditional discrete encoding in handling cyclical time characteristics (such as the numerical disconnect between December and January of the following year) and ignoring multi-scale time combination effects (such as the specific load pattern of "Friday evenings in summer"), two-dimensional sine-cosine encoding transforms the year, month, day, and hour into two-dimensional embedded vectors of closed manifolds, achieving a continuous expression of time periodic features. Simultaneously, holiday features are encoded separately and integrated into the joint date vector, effectively capturing the combination effects of different time scales. This encoding method allows the model to accurately understand the proximity relationships and scene specificity of time dimensions, avoiding the misleading judgment of periodic patterns caused by "numerical discontinuities," and laying a feature foundation for subsequent scene-based predictions.

[0017] 5. Embedding Time Distance Factor to Optimize the Weight Allocation of Time-Series Features: By embedding a time distance adjustment term based on an exponentially decaying function into the date feature combination, the model automatically assigns higher weight to recent data while focusing on global time-series information. This aligns with the principle in load forecasting that "recent data has higher reference value," and avoids the insufficient capture of recent trends caused by the indiscriminate processing of all historical data in traditional methods. This design allows the model to more sensitively perceive short-term load fluctuation trends (such as changes in electricity consumption patterns over the past 3 days) while retaining the ability to perceive key turning points in the long term (such as historical peak values ​​for the same period), thus improving the correlation between time-series features and load changes.

[0018] 6. A date-scene prediction model is constructed based on a time-delay attention mechanism. Through self-attention weights, it learns the complex interaction relationships of multi-scale date features (such as season, weekday, and hour), forming a strong correlation mapping between "scene" and "load" (e.g., "summer weekday evenings" correspond to high load, and "winter holiday daytimes" correspond to low load). This scene-based time benchmark provides a "scene anchor" for subsequent cross-lag effect analysis, avoiding misjudging inherent load patterns of scenes as the lagged effects of external driving factors. This helps the model more accurately distinguish between "scene-related fluctuations" and "driving fluctuations," improving its adaptability to load characteristics under different scenes.

[0019] 7. Extraction of Cross-Lag Effect Features and Quantification of the Lag Impact of External Drivers: Addressing the cross-lag characteristics of load and external factors (such as temperature and electricity prices) in power systems (e.g., load changes significantly only after several hours of temperature rise), a cross-lag information analysis module and a lag attention module are constructed. Batch matrix operations are used to calculate the correlations under different lag orders in parallel, avoiding the inefficiency of traditional serial calculations. Simultaneously, the lag attention module integrates original load features and lag embedding features, capturing both the load's own temporal dependence and accurately quantifying the intensity and duration of the lag impact of external driving factors (e.g., the time lag of summer temperature on load is 2-4 hours, and the time lag of winter temperature on load is 1-2 hours), solving the prediction lag or lead problem caused by traditional models ignoring time lag effects. Attached Figure Description

[0020] Figure 1 This is a flowchart of a deep learning-based short-term power load forecasting system according to an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] like Figure 1 As shown, the short-term power load forecasting system based on deep learning includes the following steps: Step s1: Perform classification sliding window settings and integrity checks on the local multi-source data uploaded by each edge node, and determine whether the local multi-source data is complete based on the integrity check results; Step s2: Perform historical same-period deviation detection on the complete data, determine whether there is any suspicious load based on the detection results, and if so, perform cross-source data cross-validation on the suspicious load. Determine whether to combine the cross-validation results with the real-time trend verification results for linkage judgment based on the cross-validation results, and remove the complete data marked as inaccurate data. Step s3: Perform two-dimensional encoding and time distance factor embedding on the complete data to generate enhanced date feature combinations. Based on the enhanced date feature combinations, perform multi-scale date feature interaction analysis and output a scenario-based time benchmark. Step s4: Extract cross-lag effect features from the numerical time series of load data and driving factors in the complete data to obtain the driving lag feature sequence. Based on the scenario-based time base and the driving lag feature sequence, obtain the load forecast value in the next load forecast time scale.

[0023] It should be further explained that in the specific implementation process, a cloud data center is built, and edge nodes are deployed in each data subject (power grid company, user, meteorological department, energy service provider) within the preset scope. The cloud data center communicates with each edge node. The edge nodes are used to de-identify the collected local multi-source data (including user historical load data, power grid operation data, meteorological data, electricity price data, holiday data, and user behavior data) and then upload it to the cloud data center. The cloud data center is equipped with several blockchain nodes, which are interconnected to form a blockchain network. Local multi-source data uploaded by edge nodes is linked to the blockchain nodes through a data uploading terminal. The data uploading terminal is used to store the local multi-source data uploaded to the cloud data center to the blockchain nodes and mark the uploading time.

[0024] It should be further explained that, in the specific implementation process, the process of classifying and sliding windowing the local multi-source data uploaded by each edge node and performing integrity checks includes: The collection period, load forecasting time scale, and weighting hierarchy of different data types in the local multi-source data of each edge node are preset (core layer (load peak period data, real-time temperature data, w=1.5), important layer (load flat period data, humidity / wind speed data, w=1.0), basic layer (load valley period data, historical electricity price data, w=0.5)), as shown in Table 1 below: Table 1

[0025] Taking 15-minute load data as an example, the basic window length is set to 1 hour (containing 4 consecutive 15-minute data points), and the sliding step is 15 minutes. That is, a new 1-hour window is generated every time one data point is moved (such as window 1:00:00-01:00, window 2:00:15-01:15, window 3:00:30-01:30, etc.). Set the base window length for different data types, and perform scenario-based calibration of the base window length for different data types based on the data body of each edge node; Scenario-based calibration includes: Critical time period definition: Critical time periods are defined according to user type (e.g., industrial users: production period from 8:00 am to 8:00 pm; residential users: peak electricity consumption periods from 7:00 am to 9:00 am and from 6:00 pm to 10:00 pm). Calibration rules: The window length for critical periods is reduced by 50% (e.g., for industrial users, the window length for 15-minute load data during production periods is changed from 1 hour to 30 minutes), while the basic window length is maintained for non-critical periods to ensure that the missing critical periods are captured more accurately.

[0026] Based on the basic window length of the scenario-based calibration, sliding windows of different data types are set for each edge node. The window missing rate of sliding windows of different data types is obtained. Based on the window missing rate of sliding windows of different data types and the weight level of different data types, the integrity score of sliding windows is obtained. ,in, For window missing rate, This represents the total number of data points within the window. This represents the number of missing data points within the window. ,in, To ensure completeness, This represents the data type, where n is the total number of data types. Let i be the window missing rate for data type i. The weight hierarchy for data type i.

[0027] A preset integrity score threshold (75%) is set. If the integrity score of the sliding window is less than the integrity score threshold, the local multi-source data in the sliding window is marked as missing data. If the integrity score of the sliding window is greater than or equal to the integrity score threshold, the local multi-source data in the sliding window is marked as complete data.

[0028] It should be further explained that, in the specific implementation process, the process of detecting historical contemporaneous deviations from complete data includes: Extract complete data from the historical periodic window of the current sliding window of the edge node from the blockchain node, mark the complete data in the historical periodic window as historical periodic data, compare the complete data of the current sliding window with the complete data of the historical periodic data (last 30 days), and obtain the load deviation rate of the load data in the complete data (load deviation rate = (current load value - average value of the same period in the last 30 days) / average value of the same period in the last 30 days × 100%) and the periodic deviation rate of other data types (periodic deviation rate = (current value - historical average) / historical average × 100%). A preset load deviation rate threshold is set. Load data with a load deviation rate greater than the load deviation rate threshold is marked as suspicious loads. Cross-source data cross-validation is performed on suspicious loads. If the suspicious load is verified to have reasonable driving factors, the suspicious load is removed from the label. If the suspicious load is found to have no reasonable driving factors, the suspicious load is marked and then combined with the real-time trend verification results (load change rate) for a joint judgment.

[0029] It should be further explained that, in the specific implementation process, the process of cross-validating suspicious loads across sources includes: Acquire data types that have potential driving logic with load data (including meteorological data, electricity price data, user behavior data, and power grid data). For example, the start-up and shutdown of industrial user production lines cause sudden load changes, and industrial users actively reduce load during peak electricity price periods. Mark the data types as driving types. Collect complete data from each blockchain node within several historical sliding windows (at least one year of historical data (covering the entire season / scenario)). Based on the driving type, extract driving features and load deviation rate from the complete data within each historical sliding window to obtain load deviation rate and driving features (temperature change amplitude, electricity price fluctuation amplitude, holidays (discrete), production plan adjustment (discrete)). Construct a feature regression model and train the feature regression model using load deviation rate and driving features as training data to obtain the trained feature regression model. In this invention, a feature regression model is constructed based on XGBoost. Through iterative fitting of multiple decision trees, the model learns to predict the load deviation rate based on the feature values. The splitting of each tree aims to reduce the prediction error of the load deviation rate.

[0030] Model training rules: Initialize the first decision tree and select the optimal splitting feature (such as temperature change) with the goal of "minimizing the prediction error of the load deviation rate"; each subsequent tree focuses on correcting the prediction error of the previous model and gradually improves the prediction accuracy of the load deviation rate. Training termination condition: The model's prediction error (such as RMSE) on the validation set no longer decreases, or the preset maximum number of trees is reached.

[0031] After the model training is completed, it is determined that when a certain driving feature is used to split a node in each decision tree, it will reduce the impurity of the dataset (measured by the mean square error of the load deviation rate). The sum of the impurity reduction of a certain driving feature in each split across all decision trees is marked as the gain score. The gain score is normalized to obtain the proportion of the load deviation rate explained by a certain driving feature. It should be further explained that for discrete characteristics (holidays, production plans, peak and off-peak periods): a scenario comparison method is used: comparing the load difference when the characteristic exists and when it does not. For example, if the current period is a statutory holiday (discrete characteristic), compare the average load of the same period "non-holiday" over the past 30 days. If the load difference rate is in the same direction as the load deviation rate of the current suspected load (e.g., both are positive deviations), and the load deviation rate is greater than the load deviation rate threshold (15%), then it is initially determined to be a strongly correlated characteristic. Subsequently, the independence between the strongly correlated characteristic and the load difference rate is verified by a chi-square test. If the test result P value < 0.05 (rejecting the "no correlation" hypothesis), then the strongly correlated characteristic is marked as a driving factor. The feature regression model that has been trained outputs the proportion of each driving type that explains the load deviation rate. The thresholds corresponding to the same period deviation rate and the explanation proportion are preset (determined based on complete historical data and expert evaluation, used to judge whether the driving factors are reliable enough). If the same period deviation rate and the explanation proportion of the driving type are both greater than the corresponding thresholds, the driving type is marked as a driving factor. By statistically analyzing the complete data of each blockchain node within several historical sliding windows, the load deviation rate range of each driving factor under different synchronous deviation rate conditions is identified, and a reasonable deviation mapping table is constructed. The reasonable deviation mapping table includes the load deviation rate range corresponding to the synchronous deviation rate range of different driving factors. For example, if the temperature increases by 8°C compared to the historical average, the corresponding load deviation rate range is ±18% to ±25%. Based on the synchronous deviation rate, explanatory proportion, and reasonable deviation mapping table of each driving factor, the total reasonable deviation rate range is obtained. Specifically, the load deviation rate range corresponding to the synchronous deviation rate of each driving factor is obtained; the weight coefficient of each driving factor is set according to the explanatory proportion of each driving factor; the load deviation rate range corresponding to the synchronous deviation rate of each driving factor is weighted and averaged according to the weight coefficient of each driving factor to obtain the total reasonable deviation rate range; it is then determined whether the load deviation rate of the suspicious load falls within the total reasonable deviation rate range. If it does, the suspicious load is determined to have a reasonable driving factor; if it does not, the suspicious load is determined to have no reasonable driving factor.

[0032] It should be further explained that, in the specific implementation process, the process of making a joint judgment based on the real-time trend verification results includes: Extract the load change rate of the current sliding window of the edge node and the historical maximum load change rate (the "physical limit fluctuation range" based on the long-term load data (at least 1 year) of the edge node, which fully includes its electricity consumption characteristics (such as the production load adjustment of industrial users, the work and rest fluctuation of residential users), extreme scenarios (historical extreme weather, sudden electricity demand), and all reasonable fluctuations, which is the "safe boundary" of the user's load change). Compare the load change rate with the historical maximum load change rate. If the load change rate is greater than the historical maximum load change rate, mark the complete data as inaccurate data, remove the inaccurate data, and require the edge node to resend the local multi-source data of the current sliding window. If the load change rate is less than or equal to the historical maximum load change rate, then extract multiple sets of historical data from the current sliding window and determine whether there is a deviation trend in the multiple sets of historical data, including: Data splitting and comparison dimensions: The nearly 90 days of data were divided into 3 groups according to a "30-day cycle" (T1: 61-90 days ago, T2: 31-60 days ago, T3: 1-30 days ago). The load value corresponding to the current time point was extracted from each group of data (e.g., if the current time is 10:00 on the 15th of each month, the load value of 10:00 on the 15th of each month is extracted from T1, T2, and T3).

[0033] Calculate the mean of each set of data for the same period: obtain the mean of T1 (L1), the mean of T2 (L2), the mean of T3 (L3), and the current load value (L0).

[0034] Trend determination rules: Increasing trend: satisfying L1<L2<L3<L0, and the deviation rate of the means of two adjacent groups ≥3% (to avoid misjudging small fluctuations as trends); Decreasing trend: satisfying L1>L2>L3>L0, and the deviation rate of the means between two adjacent groups ≥3%; No clear trend: Does not meet the above continuous increasing / decreasing pattern (e.g., L1 < L2 > L3 < L0), or the deviation rate between two adjacent groups is <3%.

[0035] If a deviation trend exists (increasing or decreasing), the historical benchmark data for the same period in the current sliding window is obtained based on the deviation trend (the original average of the same period in the past 30 days is replaced with the moving average of the average of the same period in the past 90 days (e.g., (L1+L2+L3+L0) / 4)). The complete data of the current sliding window is compared with the complete data of the historical benchmark data to obtain the load deviation rate of the load data and the same period deviation rate of other data types in the complete data. Load data with a load deviation rate greater than the load deviation rate threshold is marked as suspicious load. Cross-source data cross-validation is performed on suspicious loads. If the suspicious load is verified to have reasonable driving factors, the suspicious load is removed from the mark. If the suspicious load is verified to have no reasonable driving factors, the suspicious load is retained. Then, the real-time trend verification results are combined for linkage judgment. If there is no deviation trend, the complete data is marked as inaccurate data, the inaccurate data is removed, and the edge nodes are asked to resend the local multi-source data of the current sliding window.

[0036] It should be further explained that, in the specific implementation process, the process of performing two-dimensional encoding on the complete data and embedding time distance factors to generate enhanced date feature combinations includes: Date characteristics are an indispensable core input element in power user load forecasting. They not only carry natural periodic patterns such as daily, weekly, and annual cycles, but also include event-driven time information such as holidays and work schedule adjustments. However, the traditional discrete coding method struggles to adapt to the cyclical nature of time when representing this complex time information, easily severing the continuous correlation of time dimensions. For example, from the numerical coding results, the values ​​for December and January of the following year differ significantly, but from the actual time logic, the two are closely connected. This "numerical discontinuity" caused by discrete coding severely interferes with the model's accurate understanding of time cycle relationships.

[0037] Secondly, traditional discrete coding often treats different time scales (such as weekday attributes and seasonal attributes) as independent variables, failing to capture the complex impact of different time scale combinations on load. Taking "Friday evening in summer" as an example, the corresponding load change pattern has significant scene specificity. Using only isolated weekday or seasonal features, the model cannot effectively learn the load patterns under this combined scenario.

[0038] The load data and driving factors corresponding to the numerical time series and date data (the date data in the complete data includes four time dimensions: year, month, day, and hour) of each blockchain node within several historical sliding windows are extracted. The date data within each historical sliding window is then processed by two-dimensional encoding and transformed into multiple two-dimensional embedding vectors. These two-dimensional embedding vectors are then merged into a joint date vector. By constructing the date variable x of the historical sliding window T into a two-dimensional embedding vector, the continuous expression of periodic features is achieved. The two-dimensional encoding processing structure is as follows: x After encoding, each periodic feature forms a closed manifold on the two-dimensional plane, effectively mitigating the discontinuity problem of traditional encoding at periodic boundaries (such as December and January of the following year), and improving the handling of holiday features. The date is processed separately using a one-hot encoding method. A value of 1 indicates that the date is a holiday, and a value of 0 indicates that it is not a holiday. With a preset lag period (24h), the numerical time series corresponding to the load data within the historical sliding window and the lag period (24h after the historical sliding window) of the joint date vector are extracted. The numerical time series corresponding to the load data within the historical sliding window and the lag period are labeled as the core labels of the joint date vector. Traditional methods generally lack attention to time proximity and treat all historical time data equally, but ignore the key rule in the field of load forecasting - recent historical data often contains more valuable load change information. This indiscriminate treatment will reduce the model's accuracy in capturing recent load trends. Therefore, after the core label is labeled, the time distance factor is embedded into the joint date vector to generate an enhanced date feature combination. This enhances the weight of attention to recent time points while retaining the global modeling capability of the self-attention mechanism, so that the subsequent model can focus on strong neighboring moments and also perceive potential important turning points in the long term.

[0039] The generation of enhanced date feature combinations includes: Let the current time be t, the input length of the date feature combination prediction model be s, and the enhanced date feature combination be... ,in For the first The joint date vector at time i is used to introduce a time distance factor df between time i and time j. This df term is used to adjust the attention score between time i and time j in the subsequent self-attention mechanism. The time distance factor is: df = ,in, It is an exponentially decaying function.

[0040] It should be further explained that, in the specific implementation process, the process of performing multi-scale date feature interactive analysis based on enhanced date feature combinations to output a scenario-based time reference includes: The core labels of the enhanced date feature combinations and each joint date vector within these combinations are used as training data, with 75% (training set) and 25% (test set). A date scene prediction model is constructed based on a time-delay attention mechanism. The model is trained using the training data. After receiving the enhanced date feature combinations, the self-attention mechanism calculates the attention weights of each joint date vector with all other features (e.g., the association weights between "summer month" and "Friday" and "evening hour"), and then fuses them to obtain a context-aware vector. ,Will After being fused with the core label, the load prediction is input into the subsequent self-attention module and feedforward neural network, and outputs the load prediction value. The mean squared error (MSE) between the load prediction value and the core label is used as the loss function, and backpropagation is used to optimize the parameters of the self-attention mechanism (the weight matrix of feature interactions). Ultimately, a strong correlation is formed between date feature combinations and load values. For example, the model gradually learns that the combination of "summer (E_moy) + Friday (E_dow) + 7 pm (E_hod) + non-holiday (E_holiday)" corresponds to high load, thereby adjusting the attention weight of this combination to achieve effective modeling of complex date patterns. Based on the completed training of the date scene prediction model, the load values ​​corresponding to different date feature combinations are output, and the load values ​​corresponding to different date feature combinations are marked as scene-based time benchmarks.

[0041] The intensity and duration of the cross-hysteresis effect of external influencing factors on the load vary significantly across different scenarios. For example: "Summer + weekday daytime" scenario: Temperature has a strong cross-lag effect on load (time lag of 2-4 hours, air conditioning load rises slowly). "Winter + Nighttime" scenario: The cross-lag effect of temperature on load is weak (time lag of 1-2 hours, heating load is relatively stable). In the "holiday + daytime" scenario: the cross-lag effect of electricity prices on load almost disappears (users have no incentive to use electricity during off-peak hours).

[0042] Providing a scenario-based time reference for subsequent models allows the models to remember the inherent fluctuation patterns of load under different scenarios, providing a benchmark for identifying lagged fluctuations caused by external driving factors and avoiding misjudging the inherent fluctuations of the scenario as the lagged effects of external factors.

[0043] It should be further explained that, in the specific implementation process, the process of extracting cross-lag effect features from the load data and the corresponding numerical time series of driving factors in the complete data, obtaining the driving lag feature sequence, and obtaining the load forecast value of the target blockchain node within the next load forecast time scale based on the scenario-based time benchmark and the driving lag feature sequence includes: Obtain the load data and the corresponding numerical time series of driving factors from the complete data of the current sliding window of the target blockchain node, and extract the cross-lag effect features from the load data and the corresponding numerical time series of driving factors to obtain the driving lag feature sequence. A user load prediction model is constructed based on the RBF neural network. Complete data from several historical collection periods stored in the target blockchain node is used as training data to train the user load prediction model. The user load prediction model is trained by inputting the complete data of the current sliding window of the target blockchain node, the driving lag feature sequence, and the scenario-based time reference. Based on the user load prediction model, the load prediction value of the target blockchain node in the next load prediction time scale is output.

[0044] It should be further explained that, in the specific implementation process, the process of extracting cross-lag effect features from the numerical time series corresponding to load data and driving factors includes: Because there is a cross-lag effect between load and external influencing factors in the power system. For example, a rise in temperature does not immediately cause a surge in power load, but rather the load level will rise significantly after several hours. Therefore, a cross-lag information analysis module and a lag attention module are constructed to extract the load data and the corresponding numerical time series of driving factors from the complete data within several historical sliding windows of each blockchain node as training data. Based on the lag period, the load data and the numerical time series corresponding to the driving factors are aligned by lag time series. The aligned load data and the numerical time series corresponding to the driving factors are then input into the cross-lag information analysis module. The lag embedding matrix is ​​output from the cross-lag information analysis module. The lag information output by the cross-lag information analysis module includes: The cross-lag information analysis module takes into account the numerical time series corresponding to the driving factors as input. Numerical time series corresponding to load data ,in, Let L be the batch size, d be the sequence length, and d be the driving factor dimension. The calculation process for cross-lag information analysis is as follows: ;in, This represents the value of the i-th driving factor in batch b at time t; This represents the value of the load sequence in batch b at time t; and These are the mean values ​​of the numerical time series sequences of driving factors and load data, respectively. and τ represents the standard deviation of the numerical time series of driving factors and load data, respectively; τ is the lag coefficient, representing the time offset of load data relative to driving factors. This formula uses matrix operations to calculate the cross-lag relationship in parallel, reconstructing the traditional serial calculation logic into a parallel mode, thus avoiding the inefficiency of traditional loop calculation.

[0045] The training data and the lag embedding matrix are input into the lag attention module. The lag attention module performs a linear transformation on the training data to generate the corresponding Q vector. Then, it maps the lag embedding matrix and the training data into different K and V vectors. A scaled dot product attention mechanism is used to obtain the weights between the Q vector and different K vectors. The V vector is then weighted and summed to output the driving lag feature sequence, including: ; ; ; Among them, Q, K, The query vector, key vector, and value vector for the training data; , The key-value vector of the hysteresis embedding matrix; , These correspond to the dimensions of the query vector and the key vector, respectively. For hysteresis embedding matrix; , The weight matrix is ​​a learnable weight matrix; This is the final output of the delayed attention module. Let Q be the similarity matrix between vector Q and vector K. This is the similarity matrix between the Q vector and the key vector corresponding to the lag embedding matrix.

[0046] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A short-term power load forecasting system based on deep learning, characterized in that, Includes the following steps: Step s1: Perform classification sliding window settings and integrity checks on the local multi-source data uploaded by each edge node, and determine whether the local multi-source data is complete based on the integrity check results; Step s2: Perform historical same-period deviation detection on the complete data, determine whether there is any suspicious load based on the detection results, and if so, perform cross-source data cross-validation on the suspicious load. Determine whether to combine the cross-validation results with the real-time trend verification results for linkage judgment based on the cross-validation results, and remove the complete data marked as inaccurate data. Step s3: Perform two-dimensional encoding and time distance factor embedding on the complete data to generate enhanced date feature combinations. Based on the enhanced date feature combinations, perform multi-scale date feature interaction analysis and output a scenario-based time benchmark. Step s4: Extract cross-lag effect features from the numerical time series of load data and driving factors in the complete data to obtain the driving lag feature sequence. Based on the scenario-based time base and the driving lag feature sequence, obtain the load forecast value in the next load forecast time scale.

2. The short-term power load forecasting system based on deep learning according to claim 1, characterized in that, A cloud data center is built, and edge nodes are deployed for each data subject within a preset scope. The cloud data center communicates with each edge node. The edge nodes are used to de-identify the collected local multi-source data before uploading it to the cloud data center. The cloud data center is equipped with several blockchain nodes, which are interconnected to form a blockchain network. Local multi-source data uploaded by edge nodes is linked to the blockchain nodes through a data uploading terminal. The data uploading terminal is used to store the local multi-source data uploaded to the cloud data center to the blockchain nodes and mark the uploading time.

3. The short-term power load forecasting system based on deep learning according to claim 2, characterized in that, The process of classifying and performing integrity checks on local multi-source data uploaded from various edge nodes includes: The collection cycle, load prediction time scale, and weight level of different data types in the local multi-source data of each edge node are preset, the basic window length of different data types is set, and the basic window length of different data types is calibrated in a scenario-based manner based on the data body of each edge node. Based on the basic window length of the scenario-based calibration, sliding windows of different data types are set for each edge node. The window missing rate of sliding windows of different data types is obtained. Based on the window missing rate of sliding windows of different data types and the weight level of different data types, the integrity score of sliding windows is obtained. A preset integrity score threshold is set. If the integrity score of the sliding window is less than the integrity score threshold, the local multi-source data in the sliding window is marked as missing data. If the integrity score of the sliding window is greater than or equal to the integrity score threshold, the local multi-source data in the sliding window is marked as complete data.

4. The short-term power load forecasting system based on deep learning according to claim 3, characterized in that, The process of performing historical contemporaneous deviation detection on complete data includes: Extract complete data from the historical synchronous window of the current sliding window of the edge node from the blockchain node, mark the complete data in the historical synchronous window as historical synchronous data, compare the complete data of the current sliding window with the complete data of the historical synchronous data, and obtain the load deviation rate of the load data and the synchronous deviation rate of other data types in the complete data. A preset load deviation rate threshold is set. Load data with a load deviation rate greater than the load deviation rate threshold is marked as suspicious loads. Cross-source data cross-validation is performed on suspicious loads. If the suspicious load is verified to have reasonable driving factors, the suspicious load is removed from the label. If the suspicious load is found to have no reasonable driving factors, the suspicious load is marked and then combined with the real-time trend verification results for a joint judgment.

5. The short-term power load forecasting system based on deep learning according to claim 4, characterized in that, The process of cross-validating suspicious loads across data sources includes: Obtain the data type that has potential driving logic with the load data, and mark the data type as the driving type; Collect complete data from each blockchain node within several historical sliding windows. Based on the driving type, extract driving features and load deviation rate from the complete data within each historical sliding window to obtain load deviation rate and driving features. Construct a feature regression model and use the load deviation rate and driving features as training data to train the feature regression model to obtain the trained feature regression model. The feature regression model that has been trained outputs the proportion of each driving type that explains the load deviation rate. The thresholds corresponding to the same period deviation rate and the explanation proportion are preset. If the same period deviation rate and the explanation proportion of the driving type are both greater than the corresponding thresholds, the driving type is marked as a driving factor. A reasonable deviation mapping table is constructed, which includes the load deviation rate range corresponding to the same period deviation rate range of different driving factors. The total reasonable deviation rate range is obtained based on the same period deviation rate, explanation ratio and reasonable deviation mapping table of each driving factor. It is then determined whether the load deviation rate of the suspicious load is within the total reasonable deviation rate range. If it is, the suspicious load is determined to have reasonable driving factors. If it is not, the suspicious load is determined to have no reasonable driving factors.

6. The short-term power load forecasting system based on deep learning according to claim 5, characterized in that, The process of making a joint judgment based on real-time trend verification results includes: Extract the load change rate of the current sliding window of the edge node and the historical maximum load change rate. Compare the load change rate with the historical maximum load change rate. If the load change rate is greater than the historical maximum load change rate, mark the complete data as inaccurate data and remove the inaccurate data. If the load change rate is less than or equal to the historical maximum load change rate, then extract multiple sets of historical data from the current sliding window and determine whether there is a deviation trend in the multiple sets of historical data from the same period. If a deviation trend exists, obtain the historical benchmark data of the same period for the current sliding window based on the deviation trend, and then compare the complete data of the current sliding window with the complete data of the historical benchmark data. If there is no bias trend, then mark the complete data as inaccurate data and remove the inaccurate data.

7. The short-term power load forecasting system based on deep learning according to claim 6, characterized in that, The process of performing two-dimensional encoding on the complete data and embedding time distance factors to generate enhanced date feature combinations includes: Extract the load data and driving factors corresponding to the numerical time series and date data from the complete data of each blockchain node within several historical sliding windows. Perform two-dimensional encoding on the date data within each historical sliding window to convert it into multiple two-dimensional embedding vectors. Merge the two-dimensional embedding vectors into a joint date vector. By setting a lag period, extracting the numerical time series corresponding to the load data within the historical sliding window and lag period where the joint date vector is located, and labeling the numerical time series corresponding to the load data within the historical sliding window and lag period as the core label of the joint date vector, embedding the time distance factor into the joint date vector after completing the core label labeling, and generating an enhanced date feature combination.

8. The short-term power load forecasting system based on deep learning according to claim 7, characterized in that, The process of performing multi-scale interactive analysis of date features based on enhanced date feature combinations to output a contextualized time reference includes: The enhanced date feature combination and the core labels of each joint date vector in the enhanced date feature combination are used as training data. A date scene prediction model is constructed based on the time-delay attention mechanism. The date scene prediction model is trained using the training data. The load value corresponding to different date feature combinations is output according to the trained date scene prediction model. The load values ​​corresponding to different combinations of date features are marked as scenario-based time bases.

9. The short-term power load forecasting system based on deep learning according to claim 8, characterized in that, The process of obtaining load forecast values ​​for the next load forecast time scale includes: Obtain the load data and the corresponding numerical time series of driving factors from the complete data of the current sliding window of the target blockchain node, and extract the cross-lag effect features from the load data and the corresponding numerical time series of driving factors to obtain the driving lag feature sequence. A user load prediction model is constructed based on the RBF neural network. Complete data from several historical collection periods stored in the target blockchain node is used as training data to train the user load prediction model. The user load prediction model is trained by inputting the complete data of the current sliding window of the target blockchain node, the driving lag feature sequence, and the scenario-based time reference. Based on the user load prediction model, the load prediction value of the target blockchain node in the next load prediction time scale is output.

10. The short-term power load forecasting system based on deep learning according to claim 9, characterized in that, The process of extracting cross-lag effect features from the numerical time series corresponding to load data and driving factors includes: Construct a cross-lag information analysis module and a lag attention module, and extract the load data and the numerical time series corresponding to the driving factors from the complete data of several historical sliding windows of each blockchain node as training data; Based on the lag period, the load data and the numerical time series corresponding to the driving factors are aligned by lag time series. The aligned load data and the numerical time series corresponding to the driving factors are then input into the cross-lag information analysis module. The lag embedding matrix is ​​output from the cross-lag information analysis module. The training data and the lag embedding matrix are input into the lag attention module. The lag attention module performs a linear transformation on the training data to generate the corresponding Q vector. Then, it maps the lag embedding matrix and the training data into different K and V vectors. The scaling dot product attention mechanism is used to obtain the weights between the Q vector and different K vectors. The V vector is weighted and summed to output the driving lag feature sequence.

Images