Semantic perturbation driven load prediction residual adaptive correction method and system

Abstract

The application discloses a semantic disturbance driven load prediction residual self-adaptive correction method and system, in the offline stage, a baseline load prediction model is pre-trained by collecting structured time sequence characteristics to obtain a baseline load prediction result to generate a training set containing structured time sequence characteristics, semantic event characteristics, semantic disturbance indexes and the baseline load prediction result, which is used to train a residual gated correction module to learn the mapping relationship between semantic disturbance information and baseline load prediction residual, in the online stage, according to the corresponding structured time sequence characteristics of the time to be predicted or the period to be predicted, the final load prediction result, the prediction confidence interval and the risk level based on the prediction result are obtained by the trained residual gated correction module through the weighted gated residual. The application can maintain the stability of the baseline load prediction model in the conventional scene, when the baseline load prediction model is inaccurate due to external semantic disturbances such as extreme weather, holiday adjustment, major activities, industrial production changes, maintenance announcements, energy use notices and public events, the prediction deviation is self-adaptively corrected, and the accuracy, robustness and complex scene adaptability of the load prediction tasks such as transformer area load, regional load, park energy load, industrial load, residential load, charging load and data center power load are improved.

Description

Technical Field

[0001] This invention relates to a technology in the field of power system control, specifically a semantic disturbance-driven adaptive correction method and system for load forecast residuals. Background Technology

[0002] Existing load forecasting methods typically use structured data such as historical load sequences, meteorological data, calendar information, and operational status data as primary inputs. However, in actual operation, load time series are not only influenced by historical patterns and meteorological factors, but also by external semantic events such as policy announcements, energy consumption notices, extreme weather warnings, major social events, public emergencies, changes in industrial production arrangements, maintenance announcements, and dispatch notices. Existing semantically enhanced forecasting methods usually convert textual semantic information into numerical features and fuse them with structured features such as historical load, meteorological, and calendar data before inputting them into the forecasting model. While these methods can improve the forecasting model's ability to perceive external events to some extent, they still have the following shortcomings:

[0003] 1. The impact of semantic events on prediction results exhibits significant scene selectivity. In most conventional scenarios, prediction results are primarily determined by historical patterns, weather, and calendar factors. If semantic features are directly input into the prediction model in all scenarios, additional noise may reduce the prediction stability in conventional scenarios.

[0004] 2. A direct relationship does not always exist between semantic information and the predicted object. While some news, announcements, or policy texts may be relevant to the target region, they may not necessarily lead to significant changes in the predicted object. Directly fusing semantic features can easily result in irrelevant textual interference or over-correction.

[0005] 3. Existing methods typically use semantic features directly to predict target values, lacking explicit modeling of baseline load prediction model errors. It is difficult to determine whether semantic disturbances cause the baseline prediction to be too high or too low, and it is also difficult to determine the direction and magnitude of correction.

[0006] 4. Existing load forecasting models may have high accuracy under normal scenarios, but they are prone to inaccuracy under extreme weather, holiday adjustments, energy policy changes, industrial production changes, maintenance announcements, and emergencies. Current technology lacks a mechanism to adaptively correct the forecast residuals for abnormal scenarios without compromising the stability of the baseline load forecasting model.

[0007] 5. Existing forecasting systems often already have mature baseline load forecasting models deployed. Directly replacing them with new semantic fusion forecasting models would increase system modification costs and model validation risks. From an engineering perspective, a forecasting correction method that can be integrated into existing forecasting systems as a post-processing module is needed. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, this invention proposes a semantic perturbation-driven adaptive correction method and system for load forecast residuals. After outputting the forecast results from the baseline load forecasting model, the method utilizes semantic event information to identify potential inaccuracies in the baseline load forecast. Through candidate residual estimation and gating weight control, it maintains the stability of the baseline load forecasting model under normal scenarios. When external semantic perturbations such as extreme weather, holiday adjustments, major events, changes in industrial production, maintenance announcements, energy consumption notices, and public events cause inaccuracies in the baseline load forecasting model, the method adaptively corrects the forecasting deviations. This improves the accuracy, robustness, and adaptability to complex scenarios for load forecasting tasks such as transformer area load, regional load, park energy consumption load, industrial and commercial load, residential load, charging load, and data center power load. Furthermore, this invention can generate load forecast confidence intervals or risk levels based on the semantic perturbation index, candidate residuals, and residual correction weights, providing support for operation scheduling, load management, demand response, capacity verification, and energy optimization.

[0009] This invention is achieved through the following technical solution:

[0010] This invention relates to a semantic perturbation-driven adaptive correction method for load forecast residuals. In the offline stage, structured time-series features are collected and input into a pre-trained baseline load forecasting model to obtain baseline load forecasting results. This generates a training set containing structured time-series features, semantic event features, semantic perturbation index, and baseline load forecasting results. This training set is used to train a residual gating correction module to learn the mapping relationship between semantic perturbation information and baseline load forecasting residuals. In the online stage, based on the structured time-series features corresponding to the time or period to be predicted, the trained residual gating correction module generates a weighted gating residual to obtain the final load forecasting result, as well as the prediction confidence interval and risk level based on the forecasting result.

[0011] The structured time-series features are obtained by acquiring the structured time-series features of the target load prediction object and the textual semantic information related to the target load prediction object, specifically including:

[0012] Step 1: Obtain the target load forecast object at time [time]. The previous structured time-series features, textual semantic information related to the target load prediction object, and the actual load observations corresponding to the training phase.

[0013] The target load forecast objects include: transformer area load, regional load, park energy load, industrial and commercial load, residential load, charging load, data center power load, integrated energy park power consumption, and other load-related power time series quantities.

[0014] Step 2: Let the set of prediction step sizes be: ,in: To predict the total number of steps; when the prediction task is to predict a single future time point, =1; when the prediction task is multi-time-section prediction of the future. For the predicted sequence length. Suppose the target load prediction object is in the [number]th [sequence]. The actual load observation value of each predicted section is Structured time series features are The set of semantic events within the prediction window is: ,in: The number of semantic events related to the target load prediction object within the prediction window;

[0015] The structured time series features This includes: historical load sequences, meteorological characteristics, calendar characteristics, operational status characteristics, periodic statistical characteristics, trend characteristics, lag characteristics, peak and valley characteristics, and rate of change characteristics.

[0016] The aforementioned textual semantic information includes: news texts, policy announcement texts, weather warning texts, public event texts, holiday arrangement texts, industry operation texts, emergency texts, maintenance announcement texts, energy consumption notice texts, park announcement texts, and load management notice texts.

[0017] The baseline load prediction model is based on structured time-series features for prediction. Textual semantic information is not used as a direct input to the baseline load prediction model. It employs, but is not limited to, time series models, regression models, tree models, ensemble learning models, recurrent neural network models, temporal convolutional network models, Transformer prediction models, and multi-step temporal prediction networks.

[0018] The baseline load forecasting model is preferably a forecasting model already deployed in an existing engineering forecasting system. When training the residual gating correction module, the parameters of the baseline load forecasting model are kept frozen or not updated, and only the subsequent residual gating correction module is trained, or the baseline load forecasting model and the residual gating correction module are trained in a phased training manner.

[0019] The training specifically includes:

[0020] Step a: Input the structured time-series features into the pre-trained baseline load prediction model to obtain the baseline load prediction results. Then, construct baseline residual labels based on the difference between the actual load observations and the baseline load prediction results. This specifically includes:

[0021] a.1 Structured temporal features Input the pre-trained baseline load prediction model to obtain the first... Baseline load prediction results for each prediction section: ,in: For baseline load forecasting models, For the first Baseline load prediction results for each prediction section.

[0022] a.2 Construct baseline residual labels based on the difference between actual load observations and baseline load forecasts. ,in: For the first Baseline residual labels for each predicted section.

[0023] When baseline residual label When the baseline load forecast is lower than the actual load observation, it means that the baseline load forecast is lower than the actual load observation. When the baseline load forecast is higher than the actual load observation, it means that the baseline load forecast is higher than the actual load observation. When the value is close to 0, the baseline load forecast is relatively close to the actual load observation.

[0024] Step b: Perform semantic encoding, event recognition, time alignment, and perturbation intensity quantification on the text semantic information to obtain semantic event features and semantic perturbation index, specifically including:

[0025] b.1 to the first Semantic encoding is performed on the semantic event text to obtain the event semantic vector. ,in: For the first The text content corresponding to each semantic event. For semantic encoding functions, This is the semantic vector of the event text.

[0026] b.2 Identify the first The event type, event intensity, scope of influence, duration, and event time information of each semantic event are denoted as follows: For event type, For the intensity of the event, For the area, users, capacity, or load size affected by the event, For the duration of the event, This is the center time of the interval in which the event occurred.

[0027] Preferably, to eliminate the dimensional differences among different event attributes, the event intensity, scope of influence, duration, and time decay relationship are normalized: , , , ,in: The normalized event intensity, Event type The corresponding maximum strength calibration value; This represents the normalized range of influence. The total area, total users, total capacity, or total load scale corresponding to the target load forecast object; The normalized duration of the event. The preset maximum duration; For the first The event affects the first The time decay coefficient of each predicted section For the first The time for each predicted cross-section This is a time decay scale parameter related to the event type.

[0028] The time decay coefficient is used to characterize the change in the impact of semantic events over time. The time decay coefficient is larger when the prediction section is near the duration of the event; the time decay coefficient gradually decreases when the prediction section is far from the event occurrence period.

[0029] b.3 Construct a semantic event perturbation attribute vector containing event text semantic vector, event type embedding, normalized event intensity, normalized influence range, normalized event duration, time decay coefficient, and contextual relevance, which is used to characterize the potential perturbation effect of semantic events on the target load prediction object under a specific prediction section.

[0030] b.4 Calculate the semantic perturbation index, specifically including:

[0031] b.4.1 Map structured time-series features and baseline load forecast results to baseline state representation: ,in: This represents the baseline state. For structured timing coding functions, This is a mapping function for baseline load forecast results. To predict step-size embedding, This is the baseline state mapping function.

[0032] b.4.2 Input the semantic event perturbation attribute vector and the baseline state representation into the perturbation scoring network to obtain the semantic event perturbation latent representation. ,in: For the first The semantic event for the first The hidden representation of the perturbation of a predicted cross section. , , For trainable parameters, For nonlinear activation functions, the first... The semantic event in the ... The perturbation attribute vector of each predicted section ,in: This is a semantic event perturbation attribute vector. Event type The trainable embeddings are, symbols For vector concatenation operations, the first... The semantic events and the target load prediction object in the 1st Contextual correlation of each predicted section , and For a trainable parameter matrix, For vector dimensions, The sigmoid function is used to predict the context vector of the target load forecast object. , This is a context encoding function.

[0033] The context vector includes: prediction object type, regional attribute, time period attribute, seasonal attribute, structured time series status, baseline load prediction result, and prediction step size number information.

[0034] b.4.3 Calculate the perturbation score for a single semantic event: ,in: For the first The semantic event for the first The perturbation score of each predicted section. These are trainable parameters. The higher the perturbation score, the more likely the semantic event is to cause a bias in the baseline load prediction results.

[0035] b.4.4 When multiple semantic events exist within the same prediction window, an attention mechanism is used to aggregate the perturbation contributions of multiple events to obtain the first semantic event. The semantic perturbation index of each predicted section: ,in: For the first The semantic perturbation index of each predicted section, with a value range of [value range missing]. If no semantic event exists within the prediction window, then let .

[0036] The aforementioned aggregation, using the first The perturbation attention weights for each semantic event are: ,in: For the first The semantic event for the first The perturbation attention weights for each predicted section. These are trainable parameters.

[0037] Step c: Input the structured time-series features, semantic event features, semantic perturbation index, and baseline load prediction results into the residual gating correction module to obtain candidate residuals and residual correction weights. Then, train the residual gating correction module based on the candidate residuals, residual correction weights, baseline load prediction results, baseline residual labels, and actual load observations. This specifically includes:

[0038] c.1 Constructing a fusion representation for residual correction ,in: For the fusion representation used for residual correction, For layer normalization function, For trainable parameters, semantic event perturbation aggregation representation , For the first The semantic event perturbation aggregation representation of each prediction section. If there is no semantic event within the prediction window, then let It is a zero vector.

[0039] c.2 Based on fusion representation The residual estimation subnetwork outputs candidate residuals. ,in: For the first Candidate residuals for each predicted cross section, For the first The historical residual scaling parameter corresponding to each prediction step size can be determined by the standard deviation, mean absolute deviation or quantile scale of the historical baseline residuals. , , , } represents trainable parameters, obtained through and Function constraints allow candidate residuals to vary within a reasonable error scale, reducing the risk of overcorrection caused by anomalous semantic events.

[0040] c.3 Calculate the residual correction weights used to control the strength of the corrections made by the candidate residuals to the baseline load forecast results: ,in: For the first The residual correction weights for each predicted section, with values ​​ranging from [value range missing]. ; , , , , These are trainable parameters.

[0041] When the semantic disturbance index is low and the baseline load forecast is relatively reliable, the residual correction weight approaches 0; when the semantic disturbance index is high and the baseline forecast may have biases, the residual correction weight increases, allowing candidate residuals to participate in the correction of the final load forecast result.

[0042] c.4 Constructing a supervised loss function for the semantic perturbation index and minimizing the supervised loss, the semantic perturbation scoring network can learn from historical samples which semantic events, under what target object context, and prediction step size, are more likely to cause bias in the baseline load prediction model. Specifically: ,in: For the semantic perturbation index supervised loss, the first Soft labeling of baseline prediction inaccuracy for each prediction section , The first in the historical training samples The baseline predicted absolute residual corresponding to the prediction step size is the first... Quantiles To prevent extremely small positive numbers with a denominator of zero, To truncate the calculation results to Interval.

[0043] The closer the soft label for baseline forecast inaccuracy is to 1, the more likely the baseline load forecast is to be inaccurate; the closer it is to 0, the more reliable the baseline load forecast is.

[0044] The final load forecast results ,in: For the first Final load forecast results for each forecast section, gated residuals: , For the first Gated residuals of each predicted section.

[0045] The joint optimization training of the residual correction related network of the residual gating correction module specifically includes: during the training phase, constructing a joint training objective function: ,in: , , , The loss weighting coefficients are the final prediction error loss. Candidate residual estimation loss Gating regularization loss Residual smoothing loss , Loss function. Gated regularized loss. This is used to constrain the residual correction weights to remain small when the semantic perturbation index is low, thus avoiding overcorrection in scenarios with no or weak perturbations. Residual smoothing loss This is used to constrain the changes in gated residuals between adjacent prediction sections from excessive jumps, thereby enhancing the continuity and stability of the prediction curve.

[0046] The aforementioned prediction confidence interval is specifically as follows: Where: half width of the prediction interval , , The non-negative adjustment coefficient, the first Quantile residuals corresponding to each prediction step size , h represents the confidence level. Next Error quantiles corresponding to each prediction step size.

[0047] The risk level refers to classifying the prediction results into low-risk, medium-risk, or high-risk levels based on the relationship between the risk score and a preset threshold. When the semantic perturbation index is high, the gating residual is large, or the prediction interval is wide, the prediction risk level is increased to prompt operation scheduling, load management, or maintenance decision-making personnel to pay attention to prediction uncertainty. Specifically: ,in: For the first Predictive risk score for each prediction section, , , These are non-negative weighting coefficients.

[0048] Technical effect

[0049] This invention uses textual semantic information for residual gating correction after the baseline load prediction results, enabling the baseline load prediction model to continue to perform stable prediction functions under normal scenarios. The semantic information is mainly used for deviation correction under abnormal disturbance scenarios, thereby avoiding interference from irrelevant textual information on the normal prediction results.

[0050] This invention explicitly constructs baseline residual labels between actual load observations and baseline load prediction results, and uses these residual labels as supervisory signals to train a semantic perturbation scoring network and a residual gating correction module, enabling the model to learn the correspondence between "semantic events and baseline prediction errors." Compared to directly predicting target values, this approach more clearly reflects the impact of semantic events on the direction and magnitude of prediction bias.

[0051] The semantic perturbation index in this invention is obtained through supervised training based on semantic event attributes, the context of the target load prediction object, and the residuals of historical baseline load predictions. This mechanism avoids the problem of relying on experience to set event weights, making the perturbation intensity assessment results data-driven and trainable.

[0052] This invention constructs a semantic event perturbation attribute vector by combining event semantic vector, event type embedding, event intensity, scope of influence, duration, contextual relevance, and time decay coefficient. This vector can describe external semantic perturbations from multiple dimensions, such as event content, event attributes, event target, and event temporal relationship, thereby improving the fine-grainedness and specificity of semantic perturbation identification.

[0053] This invention uses contextual relevance calculation to match semantic events with the predicted object type, regional attributes, time period attributes, structured time series status, and baseline load prediction results. This reduces the interference of irrelevant news, announcements, or event texts on prediction correction and improves the effectiveness of utilizing semantic perturbation information.

[0054] This invention aggregates the perturbation contributions of multiple semantic events within the same prediction window through an attention mechanism. This enables the differentiation of the degree of influence of different events on different prediction sections when multiple events occur simultaneously, avoiding the problems of excessive perturbation or information dilution caused by simple summation or fixed-weight aggregation.

[0055] This invention sets up a candidate residual estimation subnetwork and a gated weight subnetwork, modeling "how much to correct" and "whether to correct" separately. The candidate residual is used to define the possible direction and magnitude of the deviation in the baseline load forecast result, and the gated weight is used to control the strength of the candidate residual's participation in the final load forecast result, thereby achieving adaptive and controllable correction of the forecast result.

[0056] This invention limits the variation range of candidate residuals by using residual scaling parameters and nonlinear constraint functions, and suppresses overcorrection in weakly disturbed scenarios and abnormal jumps between adjacent prediction sections by using gated regularization loss and residual smoothing loss, thereby improving the stability and engineering usability of the final prediction curve. Attached Figure Description

[0057] Figure 1 This is a flowchart of the present invention;

[0058] Figure 2 This is a system module diagram for an example embodiment;

[0059] Figures 3-8 This is a schematic diagram illustrating the effect of an example. Detailed Implementation

[0060] like Figure 2As shown, this embodiment relates to a semantic perturbation-driven load forecast residual gating correction system, including: a data acquisition module, a baseline load forecasting module, a residual label construction module, and a semantic event parsing module connected thereto; a residual gating correction module, a semantic perturbation index generation module, a load forecasting output module, a forecasting risk assessment module, and a model update module connected thereto; wherein: the data acquisition module is used to acquire the structured time-series features, textual semantic information, and real load observations required for the training phase of the target load forecasting object, and to perform time alignment and sample numbering on data from different sources; the baseline load forecasting module is used to generate baseline load forecasting results based on the structured time-series features; the residual label construction module is used to construct baseline residual labels based on real load observations and baseline load forecasting results; the semantic event parsing module is used to... The system performs semantic cleaning, semantic encoding, event identification, time alignment, and event attribute extraction on the text to obtain semantic event features. The semantic perturbation index generation module generates a semantic perturbation index based on semantic event features, event attributes, the target load prediction object context, and historical residual supervision signals. The residual gating correction module generates candidate residuals, residual correction weights, and gated residuals based on structured time-series features, semantic event perturbation representations, semantic perturbation indices, and baseline load prediction results. The load prediction output module overlays the gated residuals onto the baseline load prediction results to obtain the final load prediction result. The prediction risk assessment module generates prediction confidence intervals or prediction risk levels based on the semantic perturbation index, gated residuals, and historical residual distribution. The model update module continuously updates relevant model parameters based on newly added real load observations and residual samples.

[0061] The baseline load forecasting module includes a structured feature input unit, a baseline model calculation unit, and a baseline result output unit. The structured feature input unit receives historical load sequences, meteorological features, calendar features, operational status features, and time-series statistical features. The baseline model calculation unit inputs the structured time-series features into a pre-trained baseline load forecasting model to obtain the baseline load forecasting results for the target forecast section. The baseline result output unit outputs the baseline load forecasting results to the residual label construction module, the semantic perturbation index generation module, the residual gating correction module, and the load forecasting output module.

[0062] The residual gating correction module includes: an encoding projection unit, a cross-modal fusion unit, a residual estimation unit, a gating weight generation unit, and a residual correction unit. Specifically: the encoding projection unit embeds and maps structured temporal features, baseline load prediction results, semantic event perturbation representations, and semantic perturbation indices; the cross-modal fusion unit fuses structured temporal states, baseline load prediction results, and semantic perturbation information to obtain a fused representation for residual correction; the residual estimation unit outputs candidate residuals based on the fused representation; the gating weight generation unit outputs residual correction weights based on the fused representation and the semantic perturbation index; and the residual correction unit multiplies the candidate residuals by the residual correction weights to obtain the gating residuals.

[0063] like Figure 1 As shown, this embodiment illustrates the semantic perturbation-driven adaptive correction method for load forecast residuals based on the aforementioned system. In the offline phase, the data acquisition module collects the structured time-series features, textual semantic information, and actual load observations of the target load forecast object, and performs unified identification and timestamp management on the data. The baseline load forecasting module generates baseline load forecasting results based on the structured time-series features. In the training phase, the residual label construction module constructs baseline residual labels based on the actual load observations and the baseline load forecasting results. The semantic event parsing module performs event recognition and attribute extraction on the textual semantic information. The semantic perturbation index generation module generates a semantic perturbation index based on semantic event attributes and historical residual supervision signals. The residual gating correction module... The above information is used to train the candidate residual estimation and gating correction capabilities. During the prediction phase, the semantic event parsing module and the semantic perturbation index generation module generate a semantic perturbation index based on the textual semantic information within the current prediction window. The residual gating correction module outputs candidate residuals and residual correction weights based on structured time-series features, baseline load prediction results, and semantic perturbation information. The load prediction output module superimposes the gating residuals onto the baseline load prediction results to obtain the final load prediction results. The prediction risk assessment module further outputs the load prediction confidence interval or prediction risk level. The model update module updates the residual sample library after the actual load observations are transmitted back, and performs rolling optimization of the residual gating correction module and prediction risk assessment parameters as needed.

[0064] Through specific practical experiments, load forecast samples were constructed using publicly available half-hour load demand data. Historical load, calendar features, and time-series statistical features were used as structured inputs, and bank holidays, representative weather disturbance days, and public event dates were used as event feature inputs after being processed by the semantic event parsing module. This formed a verification process of "structured baseline forecasting - semantic disturbance index generation - candidate residual estimation - gated residual correction - final load forecast output".

[0065] This experiment verifies the effectiveness of the invention from three levels: First, it verifies whether the final load prediction error is lower than that of the structured baseline model; second, it verifies whether the error improvement on semantic disturbance days is more significant; and third, it verifies whether the gating weights can be increased during disturbance periods and relatively suppressed during normal periods, thereby avoiding excessive interference of semantic information on normal prediction results.

[0066] This experiment selected half-hour load demand data from 2020 to 2022 as the measured load sequence. After timestamp parsing, duplicate record removal, missing value checking, and time sequence arrangement, the raw data formed a total of 52,608 half-hour load sections. After constructing lagging load, rolling statistical features, and event features, the model sample retained a total of 52,272 valid samples.

[0067] The forecasting task was set to half-hourly load forecasting, meaning that for each forecast section, the load forecast value for that section was output based on historical load, calendar, lag statistics, and semantic perturbation information. To facilitate the illustration of the daily-scale operation effect, the experimental results were further aggregated by natural day to form a representative daily forecast curve, a daily-scale error curve, and a normal daily / perturbation daily error distribution.

[0068] The structured time-series features used in this experiment include: month, day of the week, hour, half-hour section number, weekday / weekend identifier, historical load of the same section over 1 day, historical load of the same section over 2 days, historical load of the same section over 7 days, previous day's average, previous day's maximum, previous day's minimum, previous day's peak-to-valley difference, and several rolling statistical features. These features are used to train the structured baseline load forecasting model.

[0069] In this experiment, semantic event features include: bank holidays, weather disturbance proxy events, and public event dates. Specifically, bank holidays characterize changes in public work and rest schedules and electricity consumption behavior; weather disturbance proxy events characterize the potential impact of abnormal weather such as storms, high temperatures, and cold waves on load; and public events characterize the disturbances to the load curve caused by social activities, public policies, or major events. In this experiment, these semantic events are equivalent to the structured event results output by the text semantic parsing module, including: event type, event name, event date, and semantic disturbance intensity identifier.

[0070] The valid samples are divided into training, validation, and test sets in chronological order. Specifically: the training set comprises 30,336 samples from January 2020 to September 2021; the validation set comprises 4,416 samples from October to December 2021; and the entire year of 2022 serves as the test set, containing 17,520 samples. This division method avoids future information leakage and meets the time-series validation requirements for load forecasting tasks.

[0071] To verify the effectiveness of the present invention, three sets of comparative models were set up: M1 is a structured baseline model, which uses only historical load, calendar, and time-series statistical features for prediction, representing conventional load prediction methods that do not use semantic perturbation residual correction. M2 is a semantic direct fusion model, which directly concatenates semantic event features with structured features and inputs them into the prediction model, used to compare with methods that "directly use semantic features". M3 is a gateless residual correction model, which estimates candidate residuals based on the baseline load prediction results, but does not use gating weights to control the correction strength, used to verify the necessity of the gating mechanism. All four sets of models use the same training set, validation set, and test set to ensure that the differences in prediction results mainly come from differences in semantic perturbation modeling methods and residual correction structures.

[0072] The output of this experiment includes five categories of results: 1) baseline load forecast results; 2) semantic perturbation index; 3) candidate residuals; 4) residual correction weights and gated residuals; and 5) final load forecast results and their error evaluation results. Mean absolute error (MAE), root mean square error (RMSE), and symmetric mean absolute percentage error (SMAPE) are used as overall forecast accuracy indicators. Furthermore, the test days are divided into normal days and semantic perturbation days according to the presence or absence of semantic events. The MAE for normal days and perturbation days are calculated separately to verify the correction effect of this invention under abnormal semantic perturbation scenarios.

[0073] Table 1 shows the overall prediction results of the four models on the 2022 test set. It can be seen that the M4 model of this invention achieves the best results across all five indices: MAE, RMSE, SMAPE, MAE on normal days, and MAE on disturbed days.

[0074] Table 1

[0075] As shown in Table 1, the overall MAE of the present invention in M4 is 1159.99MW, which is lower than 1194.87MW in the structured baseline model of M1, representing a decrease of 2.92% in overall MAE; RMSE decreases from 1650.35MW to 1599.94MW, a decrease of 3.05%; and SMAPE decreases from 4.54% to 4.42%, a decrease of 2.74%.

[0076] Furthermore, the MAE of the M4 model on semantically perturbed days decreased from 1605.91MW of the M1 model to 1401.90MW, a reduction of 12.70%; the MAE on normal days decreased from 1147.11MW to 1131.88MW, a reduction of 1.33%. This demonstrates that the present invention not only improves the overall prediction accuracy but also shows a more significant improvement on semantically perturbed days, meeting the design goal of "strengthening and correcting abnormal perturbed scenarios while maintaining stability in normal scenarios".

[0077] Meanwhile, the M4 model also achieved a lower error compared to the M3 ungated residual model, indicating that residual estimation alone is not enough to guarantee the prediction effect. Gating weights are still needed to control the intensity of residual correction in order to reduce the risk of overcorrection.

[0078] like Figure 3 As shown, September 19, 2022, was selected as a representative semantic disturbance day to compare the actual load curve, the structured baseline prediction curve, the semantic direct fusion prediction curve, and the prediction curve of this invention. This date contains obvious semantic event factors, making it suitable for testing the model's adaptability to external disturbance scenarios. As can be seen from the figure, the structured baseline model exhibits biases during some load peaks and curve transition periods, indicating that relying solely on historical load and calendar statistical characteristics is insufficient for characterizing event-driven load pattern changes. This invention, by superimposing gated residuals on the baseline prediction, produces a prediction curve that is generally closer to the actual load curve, especially demonstrating better tracking ability during periods of rapid load change and peak-valley transitions.

[0079] like Figure 4 The figure shows the changes in the intraday semantic perturbation index, residual correction weights, and baseline error normalization results. This figure is used to verify whether the semantic perturbation index and gating weights can respond consistently to the baseline prediction error. As can be seen from the figure, during periods of strong semantic perturbation, the semantic perturbation index remains at a high level, and the residual correction weights increase accordingly, indicating that the system judges a high risk of inaccuracy in the baseline prediction during this period and allows more candidate residuals to participate in the correction of the final load prediction result. During periods of weak semantic perturbation or small baseline prediction error, the gating weights are relatively limited to avoid unnecessary corrections to the baseline load prediction result by candidate residuals. This result verifies the effectiveness of the "semantic perturbation identification—gating weight control" method of this invention.

[0080] like Figure 5 The figure illustrates the relationship between baseline residuals, candidate residuals, and gated residuals. Baseline residuals reflect the deviation between the actual load and the baseline load forecast; candidate residuals reflect the correction direction and magnitude of the residual estimation unit output; and gated residuals are the actual correction amount after gating weights based on the candidate residuals. As can be seen from the figure, candidate residuals can effectively reflect the changing trend of baseline errors, but their correction magnitude is not directly added entirely to the baseline load forecast result; instead, it is formed as gated residuals after gating weights. This mechanism can enhance correction during periods of large errors and suppress correction during periods of small errors or weak disturbances, thereby improving forecast stability. This result demonstrates that the residual gating correction of this invention is not simple error compensation, but rather an adaptive correction process with semantic disturbance awareness and correction strength constraints.

[0081] like Figure 6The figure shows the daily-scale MAE variation curves of the structured baseline model and the present invention during the 2022 testing period. This figure is used to verify whether the error improvement of the present invention is sustainable, rather than just effective on individual dates. As can be seen from the figure, the daily-scale error of the present invention is lower than or close to that of the structured baseline model on most days during the testing period, especially near dates with semantic perturbations, where the error reduction is more significant. This result indicates that the semantic perturbation residual gating correction module can work stably over a long testing period and has a stronger error suppression capability for perturbation dates.

[0082] like Figure 7 The figure shows the error distribution of different models under normal days and semantically disturbed days. This figure is used to verify whether the present invention can improve the prediction accuracy on disturbed days without significantly compromising the prediction stability on normal days. As can be seen from the figure, the error distribution on semantically disturbed days is generally higher than that on normal days, indicating that external events do indeed increase the difficulty of load prediction. Compared with the structured baseline model, the error distribution of the present invention is significantly narrowed on semantically disturbed days, with both the median error and high-error samples decreasing; on normal days, the present invention maintains an error level similar to or slightly better than the baseline model, indicating that the gating mechanism can effectively suppress over-correction in undisturbed scenarios.

[0083] In addition to point prediction results, this application example can also generate prediction intervals or risk warnings based on semantic perturbation index, gated residual magnitude, and historical residual distribution. For example... Figure 8 The diagram illustrates the daily forecast interval and risk output. This section serves as an extended application of the invention in the specification, showcasing its practical value in load management and operational decision support. As the diagram shows, the forecast interval widens during periods of strong semantic disturbances or large gated residuals, indicating that the system can alert to increased uncertainty in the current load forecast. This output not only provides the final load forecast value but also offers risk awareness information to operators, facilitating advance planning of backup, demand response, or load management measures.

[0084] Experimental results show that the present invention outperforms the structured baseline model in terms of overall prediction accuracy, prediction accuracy on perturbation days, and error stability. In particular, on semantic perturbation days, the MAE of the present invention is reduced by 12.70% compared to the structured baseline model, indicating that the semantic perturbation index and the gated residual correction mechanism can effectively improve the robustness of load prediction under abnormal event scenarios.

[0085] Meanwhile, the present invention did not show significant error degradation on normal days, indicating that the gating weight can suppress the interference of irrelevant semantic information on the baseline load prediction results, enabling the system to have the technical effect of "maintaining stability in normal scenarios and adaptive correction in disturbed scenarios".

[0086] Furthermore, this invention can be integrated as a post-processing correction module into existing load forecasting systems without completely replacing the original baseline load forecasting model. This reduces engineering modification costs and model migration risks. It is applicable to various load forecasting scenarios, including transformer area load, regional load, park energy load, industrial and commercial load, residential load, charging load, and data center power load. It can also generate prediction confidence intervals or risk levels based on semantic disturbance index, gated residuals, and historical residual distribution. In addition to outputting the final prediction value, it can also indicate the uncertainty level of the prediction results, providing auxiliary support for operation scheduling, load management, standby configuration, load management, and operation and maintenance decisions.

[0087] In summary, this invention can perform residual gating correction on the baseline load forecasting results through semantic perturbation information without replacing the baseline load forecasting model. It is applicable to various load forecasting scenarios, such as transformer area load, regional load, park energy load, residential load, industrial and commercial load, charging load, and data center power load.

[0088] 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 perturbation-driven adaptive correction method for load forecast residuals, characterized in that, In the offline phase, structured time-series features are collected and input into the pre-trained baseline load prediction model to obtain the baseline load prediction results. This generates a training set containing structured time-series features, semantic event features, semantic disturbance index, and baseline load prediction results. This training set is used to train the residual gating correction module to learn the mapping relationship between semantic disturbance information and baseline load prediction residuals. In the online phase, based on the structured time-series features corresponding to the time to be predicted or the period to be predicted, the trained residual gating correction module generates a weighted gating residual to obtain the final load prediction result, as well as the prediction confidence interval and risk level based on the prediction result. The structured time series features are obtained by acquiring the structured time series features of the target load prediction object and the textual semantic information related to the target load prediction object; The baseline load prediction model is based on structured time-series features for prediction. Textual semantic information is not used as a direct input to the baseline load prediction model. When training the residual gating correction module, the parameters of the baseline load prediction model are frozen or not updated. Only the subsequent residual gating correction module is trained, or the baseline load prediction model and the residual gating correction module are trained in a phased training manner.

2. The semantic perturbation-driven adaptive correction method for load forecast residuals according to claim 1, characterized in that, The structured time series features are obtained in the following ways: Step 1: Obtain the target load forecast object at time [time]. The previous structured time-series features, textual semantic information related to the target load prediction object, and the actual load observations corresponding to the training phase; The target load forecast objects include: transformer area load, regional load, park energy load, industrial and commercial load, residential load, charging load, data center power load, integrated energy park power consumption, and other load-related power time series quantities. Step 2: Let the set of prediction step sizes be: ,in: To predict the total number of steps; when the prediction task is to predict a single future time point, =1; when the prediction task is multi-time-section prediction of the future. To predict the sequence length, let the target load prediction object be in the _th ... The actual load observation value of each predicted section is Structured time series features are The set of semantic events within the prediction window is: ,in: The number of semantic events related to the target load prediction object within the prediction window; The structured time-series features This includes: historical load series, meteorological characteristics, calendar characteristics, operational status characteristics, periodic statistical characteristics, trend characteristics, lag characteristics, peak and valley characteristics, and rate of change characteristics; The aforementioned textual semantic information includes: news texts, policy announcement texts, weather warning texts, public event texts, holiday arrangement texts, industry operation texts, emergency texts, maintenance announcement texts, energy consumption notice texts, park announcement texts, and load management notice texts.

3. The semantic perturbation-driven adaptive correction method for load forecast residuals according to claim 1, characterized in that, The training specifically includes: Step a: Input the structured time series features into the pre-trained baseline load prediction model, obtain the baseline load prediction results, and construct the baseline residual labels based on the difference between the actual load observations and the baseline load prediction results. Step b: Perform semantic encoding, event recognition, time alignment, and perturbation intensity quantification on the text semantic information to obtain semantic event features and semantic perturbation index; Step c: Input the structured time series features, semantic event features, semantic disturbance index and baseline load prediction results into the residual gating correction module to obtain candidate residuals and residual correction weights. Then, train the residual gating correction module based on the candidate residuals, residual correction weights, baseline load prediction results, baseline residual labels and actual load observations.

4. The semantic perturbation-driven adaptive correction method for load forecast residuals according to claim 3, characterized in that, Step a specifically includes: a.1 Structured temporal features Input the pre-trained baseline load prediction model to obtain the first... Baseline load prediction results for each prediction section: ,in: For baseline load forecasting models, For the first Baseline load forecast results for each forecast section; a.2 Construct baseline residual labels based on the difference between actual load observations and baseline load forecasts. ,in: For the first Baseline residual labels for each predicted section; When baseline residual label When the baseline load forecast is lower than the actual load observation, it means that the baseline load forecast is lower than the actual load observation. When the baseline load forecast is higher than the actual load observation, it means that the baseline load forecast is higher than the actual load observation. When the value is close to 0, the baseline load forecast is relatively close to the actual load observation.

5. The semantic perturbation-driven adaptive correction method for load forecast residuals according to claim 4, characterized in that, To eliminate the dimensional differences among different event attributes, the event intensity, scope of influence, duration, and time decay relationship are normalized: , , , ,in: The normalized event intensity, Event type The corresponding maximum strength calibration value; This represents the normalized range of influence. The total area, total users, total capacity, or total load scale corresponding to the target load forecast object; The normalized duration of the event. The preset maximum duration; For the first The event affects the first The time decay coefficient of each predicted section For the first The time for predicting a cross-section The time decay scale parameter is related to the event type; The time decay coefficient is used to characterize the change of the influence of semantic events over time. When the prediction section is near the duration of the event, the time decay coefficient is large; when the prediction section is far away from the time period of the event, the time decay coefficient gradually decreases.

6. The semantic perturbation-driven adaptive correction method for load forecast residuals according to claim 3, characterized in that, Step b specifically includes: b.1 to the first Semantic encoding is performed on the semantic event text to obtain the event semantic vector. ,in: For the first The text content corresponding to each semantic event. For semantic encoding functions, For event text semantic vectors; b.2 Identify the first The event type, event intensity, scope of influence, duration, and event time information of each semantic event are denoted as follows: For event type, For the intensity of the event, For the area, users, capacity, or load size affected by the event, For the duration of the event, The center time of the interval in which the event occurred; b.3 Construct a semantic event perturbation attribute vector containing event text semantic vector, event type embedding, normalized event intensity, normalized influence range, normalized event duration, time decay coefficient, and contextual relevance, which is used to characterize the potential perturbation effect of semantic events on the target load prediction object under a specific prediction section. b.4 Calculate the semantic perturbation index.

7. The semantic perturbation-driven adaptive correction method for load forecast residuals according to claim 6, characterized in that, Step b4 specifically includes: b.4.1 Map structured time-series features and baseline load forecast results to baseline state representation: ,in: This represents the baseline state. For structured timing coding functions, This is a mapping function for baseline load forecast results. To predict step-size embedding, Baseline state mapping function; b.4.2 Input the semantic event perturbation attribute vector and the baseline state representation into the perturbation scoring network to obtain the semantic event perturbation latent representation. ,in: For the first The semantic event for the first The hidden representation of the perturbation of a predicted cross section. , , For trainable parameters, For nonlinear activation functions, the first... The semantic event in the ... The perturbation attribute vector of each predicted section ,in: This is a semantic event perturbation attribute vector. Event type The trainable embeddings are, symbols For vector concatenation operations, the first... The semantic events and the target load prediction object in the 1st Contextual correlation of each predicted section , and For a trainable parameter matrix, For vector dimensions, The sigmoid function is used to predict the context vector of the target load forecast object. , For context encoding functions; The context vector includes: prediction object type, regional attribute, time period attribute, seasonal attribute, structured time series status, baseline load prediction result, and prediction step size number information; b.4.3 Calculate the perturbation score for a single semantic event: ,in: For the first The semantic event for the first The perturbation score of each predicted section. For trainable parameters, the higher the perturbation score, the more likely the semantic event will cause a deviation in the baseline load prediction results; b.4.4 When multiple semantic events exist within the same prediction window, an attention mechanism is used to aggregate the perturbation contributions of multiple events to obtain the first semantic event. The semantic perturbation index of each predicted section: ,in: For the first The semantic perturbation index of each predicted section, with a value range of [value range missing]. If no semantic event exists within the prediction window, then let ; The aforementioned aggregation, using the first The perturbation attention weights for each semantic event are: ,in: For the first The semantic event for the first The perturbation attention weights for each predicted section. These are trainable parameters.

8. The semantic perturbation-driven adaptive correction method for load forecast residuals according to claim 3, characterized in that, Step c specifically includes: c.1 Constructing a fusion representation for residual correction ,in: For the fusion representation used for residual correction, For layer normalization function, For trainable parameters, semantic event perturbation aggregation representation , For the first The semantic event perturbation aggregation representation of each prediction section, if there is no semantic event within the prediction window, then let It is a zero vector; c.2 Based on fusion representation The residual estimation subnetwork outputs candidate residuals. ,in: For the first Candidate residuals for each predicted cross section, For the first The historical residual scaling parameters corresponding to each prediction step size are determined by the standard deviation, mean absolute deviation or quantile scale of the historical baseline residuals. , , , } represents trainable parameters, obtained through and Function constraints allow candidate residuals to vary within a reasonable error scale, reducing the risk of overcorrection caused by anomalous semantic events; c.3 Calculate the residual correction weights used to control the strength of the corrections made by the candidate residuals to the baseline load forecast results: ,in: For the first The residual correction weights for each predicted section, with values ​​ranging from [value range missing]. ; , , , , These are trainable parameters; When the semantic disturbance index is low and the baseline load forecast is relatively reliable, the residual correction weight approaches 0; when the semantic disturbance index is high and the baseline forecast may have biases, the residual correction weight increases, allowing candidate residuals to participate in the correction of the final load forecast result. c.4 Constructing the supervised loss function for the semantic perturbation index and minimizing the supervised loss, the semantic perturbation scoring network can learn from historical samples which semantic events, target object contexts, and prediction step sizes are more likely to cause bias in the baseline load prediction model. Specifically: ,in: For the semantic perturbation index supervised loss, the first Soft labeling of baseline prediction inaccuracy for each prediction section , The first in the historical training samples The baseline predicted absolute residual corresponding to the prediction step size is the first... Quantiles To prevent extremely small positive numbers with a denominator of zero, To truncate the calculation results to interval; The closer the soft label for baseline forecast inaccuracy is to 1, the more likely the baseline load forecast is to be inaccurate; the closer it is to 0, the more reliable the baseline load forecast is. The final load forecast results ,in: For the first Final load forecast results for each forecast section, gated residuals: , For the first Gated residuals of each predicted section; The joint optimization training of the residual correction related network of the residual gating correction module specifically includes: during the training phase, constructing a joint training objective function: ,in: , , , The loss weighting coefficients represent the final prediction error loss. Candidate residual estimation loss Gating regularization loss Residual smoothing loss , Loss function, gated regularized loss This constraint keeps the residual correction weights small when the semantic perturbation index is low, thus avoiding overcorrection in scenarios with no or weak perturbations. (Residual smoothing loss) This is used to constrain the changes in gated residuals between adjacent prediction sections from excessive jumps, thereby enhancing the continuity and stability of the prediction curve. The aforementioned prediction confidence interval is specifically as follows: Where: half width of the prediction interval , , The non-negative adjustment coefficient, the first Quantile residuals corresponding to each prediction step size , h represents the confidence level. Next Error quantiles corresponding to each prediction step size; The aforementioned risk level refers to classifying prediction results into low-risk, medium-risk, or high-risk levels based on the relationship between the risk score and a preset threshold. When the semantic perturbation index is high, the gating residual is large, or the prediction interval is wide, the prediction risk level is increased to prompt operation scheduling, load management, or maintenance decision-making personnel to pay attention to prediction uncertainty. Specifically: ,in: For the first Predictive risk score for each prediction section, , , These are non-negative weighting coefficients.

9. A semantically perturbation-driven adaptive correction system for load forecast residuals that implements the method of any one of claims 1-8, characterized in that, include: The system comprises a data acquisition module, a baseline load prediction module, a residual label construction module, a semantic event parsing module, a residual gating correction module, a semantic perturbation index generation module, a load prediction output module, a prediction risk assessment module, and a model update module, all connected to the residual gating correction module. Specifically: the data acquisition module acquires the structured time-series features, textual semantic information, and real load observations required for training of the target load prediction object, and performs time alignment and sample numbering on data from different sources; the baseline load prediction module generates baseline load prediction results based on the structured time-series features; the residual label construction module constructs baseline residual labels based on real load observations and baseline load prediction results; and the semantic event parsing module cleans and semantically encodes the textual semantic information. The system comprises several modules: event identification, time alignment, and event attribute extraction to obtain semantic event features; a semantic perturbation index generation module to generate a semantic perturbation index based on semantic event features, event attributes, target load prediction object context, and historical residual supervision signals; a residual gating correction module to generate candidate residuals, residual correction weights, and gated residuals based on structured time-series features, semantic event perturbation representations, semantic perturbation indices, and baseline load prediction results; a load prediction output module to superimpose the gated residuals onto the baseline load prediction results to obtain the final load prediction results; a prediction risk assessment module to generate prediction confidence intervals or prediction risk levels based on semantic perturbation indices, gated residuals, and historical residual distributions; and a model update module to continuously update relevant model parameters based on newly added real load observations and residual samples.

10. The semantic perturbation-driven adaptive correction system for load forecast residuals according to claim 9, characterized in that, The baseline load forecasting module includes a structured feature input unit, a baseline model calculation unit, and a baseline result output unit. The structured feature input unit receives historical load sequences, meteorological features, calendar features, operational status features, and time-series statistical features. The baseline model calculation unit inputs the structured time-series features into a pre-trained baseline load forecasting model to obtain the baseline load forecasting results for the target forecast section. The baseline result output unit outputs the baseline load forecasting results to the residual label construction module, the semantic perturbation index generation module, the residual gating correction module, and the load forecasting output module. The residual gating correction module includes: an encoding projection unit, a cross-modal fusion unit, a residual estimation unit, a gating weight generation unit, and a residual correction unit. Specifically: the encoding projection unit embeds and maps structured temporal features, baseline load prediction results, semantic event perturbation representations, and semantic perturbation indices; the cross-modal fusion unit fuses structured temporal states, baseline load prediction results, and semantic perturbation information to obtain a fused representation for residual correction; the residual estimation unit outputs candidate residuals based on the fused representation; the gating weight generation unit outputs residual correction weights based on the fused representation and the semantic perturbation index; and the residual correction unit multiplies the candidate residuals by the residual correction weights to obtain the gating residuals.

Images