A method for load elasticity analysis based on a large language model

By extracting business event information using a large language model and combining it with an LSTM model, filtering similar historical event days, and calculating load elasticity correction, the problem of large load forecasting errors in existing technologies is solved, achieving higher forecasting accuracy and stability.

CN123121491APending Publication Date: 2026-11-17SICHUAN HUADIAN ELECTRIC POWER INVESTMENT CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611245402.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing load forecasting methods struggle to utilize business notification texts, traditional similarity day methods are prone to mismatches of business scenarios, and the LSTM-based model exhibits significant prediction errors on business event days.

Method used

The event type, direction of impact, intensity of impact, and duration of effect in business notifications are extracted using a large language model. Combined with the basic prediction values ​​of the Long Short-Term Memory Network, the prediction results of the LSTM model are corrected by screening similar historical event days and calculating load elasticity correction.

Benefits of technology

It improved the stability and accuracy of load forecasting, especially on business event days, reducing forecasting errors, with MAPE decreasing from 4.82% to 1.92%.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121491A_ABST
    Figure CN123121491A_ABST
Patent Text Reader

Abstract

This invention discloses a load elasticity analysis method based on a large language model, belonging to the field of power load forecasting and power system data analysis technology. It acquires historical load data, meteorological and calendar attributes, and business notification texts from electricity sales companies, inputs this data into a Long Short-Term Memory (LSTM) network, and outputs the basic load forecast value for the next 24 hours. The business notification text is then input into a large language model to extract structured event semantic features such as event type, direction of influence, intensity of influence, start time, and end time, without directly generating load forecast values. When a business event exists on the forecast day, it first filters similar historical event days by event type, then compares the direction of influence, intensity of influence, duration of action, and distance from traditional features to select historically similar days. The load elasticity correction amount is obtained by weighting the prediction residuals of historically similar days according to their similarity, and this corrects the basic load forecast value. By utilizing natural language business notifications, it reduces mismatches in business scenarios and improves the prediction accuracy for business event days.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the fields of power load forecasting, artificial intelligence and power system data analysis technology, and specifically relates to a load elasticity analysis method based on a large language model. Background Technology

[0002] With the ongoing reform of the electricity market, electricity sales companies need to formulate electricity purchase and sales plans based on the future electricity demand of their customers. Load forecasting errors affect electricity consumption declarations, purchase costs, and deviation assessments; therefore, improving the accuracy of short-term load forecasting by electricity sales companies is of great significance.

[0003] In addition to being affected by conventional factors such as historical load, temperature, humidity, weekdays, and holidays, the aggregated load of electricity sales companies is also affected by business events such as increased or decreased production by enterprises, equipment maintenance, temporary shutdowns, and demand response. These business events are usually communicated in the form of natural language notifications and may cause a significant increase or decrease in load during specific periods.

[0004] Existing deep learning models such as LSTM can learn the periodic and nonlinear variation patterns of historical loads, but their inputs are usually numerical variables, making it difficult to directly utilize natural language business notifications. The basic prediction model is prone to significant overestimation or underestimation on days with sudden business events.

[0005] Traditional similarity date methods typically select historical dates based on characteristics such as temperature, humidity, weekday attributes, holiday attributes, and historical workload. When the predicted date and historical dates have similar general characteristics but different business event types, traditional methods may select historical dates with inconsistent business scenarios, leading to inaccurate residual correction results.

[0006] Large language models possess strong natural language understanding and information extraction capabilities, enabling them to extract event types, impact directions, impact strengths, and durations from business notifications. However, directly generating continuous load values ​​from large language models suffers from insufficient stability and repeatability. Summary of the Invention

[0007] The technical problem this invention aims to solve is: addressing the issues that existing load forecasting methods struggle to utilize business notification texts, traditional similar-day methods are prone to mismatches of business scenarios, and LSTM-based models have significant prediction errors on business event days. This invention provides a load resilience analysis method based on a large language model.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: Firstly, a load resilience analysis method based on a large language model is provided, including the following steps: S1, acquire historical load data, meteorological data, calendar attribute data and business notification text from the electricity sales company, input the historical load data, meteorological data and calendar attribute data into the long short-term memory network, and output the basic load forecast value for the next 24 hours of the forecast day; S2, input the business notification text into the large language model, and extract only the structured event semantic features of event type, direction of influence, intensity of influence, start time and end time. The large language model does not directly generate load prediction values. S3, when there is a business event on the predicted date, first filter the same type of historical event days according to the event type, then compare the direction of influence, the intensity of influence, the start time, the end time and the distance of traditional features, and select historically similar days; S4. Calculate the prediction residuals for each of the historical similar days, weight the load elasticity correction amount according to the similarity between the historical similar days and the prediction days, and add the load elasticity correction amount to the base load prediction value to obtain the final load prediction value.

[0009] Specifically, in step S2, the event type is represented by one-hot encoding, and the influence direction and influence intensity are combined into a directional influence intensity. The directional influence intensity, together with the start time and the end time, constitutes an 8-dimensional event semantic feature. When there are no business events on the prediction day, the event type encoding, influence intensity, and duration of action are all processed as if there were no events.

[0010] Specifically, in step S3, the traditional feature distance is calculated based on the traditional similar day feature vector. The traditional similar day feature vector includes the average temperature, average humidity, weekday indicator, holiday indicator, weekday attribute, and average load of the previous day for the predicted day, expressed as: The traditional feature distance is: ; Where, in the formula, To predict the daily average temperature, To predict the daily average humidity, As a sign of a workday As a sign for holidays, For weekdays, To predict the average load of the previous day, and These are the standardized traditional similar day features of the predicted day and the historical candidate day, respectively, where j=1,2,…,6.

[0011] Specifically, in step S3, the number of historical similar days is 5; when there is no business event on the predicted day, a historical similar day is selected from the historical dates without business events.

[0012] Specifically, in step S4, the prediction residual for time h on the k-th historical similar day is: ,in, For the actual load on historically similar days, The LSTM is used to predict the load; when the residual is greater than 0, it indicates that the Long Short-Term Memory (LSTM) network underestimates the load, and when the residual is less than 0, it indicates that the LSTM network overestimates the load.

[0013] Specifically, in step S4, the load elasticity correction for time h on day d is predicted as follows: ; in, For the set of historically similar days for the predicted date, For the first Normalized weights corresponding to historically similar dates For the similar day The prediction residual at each time point; The final load forecast value at time h on day d is: ,in, To predict the final load forecast value at time h on day d; This represents the base prediction value of the LSTM at the corresponding time.

[0014] Specifically, historical load data, meteorological data, and calendar attribute data are time-aligned, missing data are filled in, and continuous variables are normalized according to the training set parameters; the data of the 168 consecutive hours before the prediction date is used as an input sample, and the load of the next 24 hours before the prediction date is used as the prediction target.

[0015] Specifically, the Long Short-Term Memory network has 5 input features, 64 hidden units, a batch size of 32, an initial learning rate of 0.001, and a maximum number of training rounds of 100.

[0016] Secondly, a load elasticity analysis system based on a large language model is provided, comprising: a data acquisition module for acquiring historical load data, meteorological data, calendar attribute data, and business notification text from the power sales company; a basic forecast module for inputting historical load data, meteorological data, and calendar attribute data into a long short-term memory network and outputting the basic load forecast value for the next 24 hours of the forecast date; an event semantic extraction module for inputting the business notification text into the large language model and extracting only the structured event semantic features of event type, influence direction, influence intensity, start time, and end time; the large language model does not directly generate load forecast values; a similar day selection module for selecting historical similar days when there are business events on the forecast date, first filtering similar historical event days according to event type, then comparing the influence direction, influence intensity, start time, end time, and distance with traditional features; and a correction module for calculating the forecast residuals of each historical similar day, weighting the load elasticity correction amount according to the similarity between the historical similar day and the forecast day, and superimposing the load elasticity correction amount onto the basic load forecast value to obtain the final load forecast value.

[0017] Thirdly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the above-mentioned load resilience analysis method based on a large language model.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) Use a large language model to convert natural language business notifications into structured event semantic features, so that the load forecasting model can utilize text information such as enterprise production adjustments, equipment maintenance, temporary shutdowns and demand response; (2) The large language model is only used for event information extraction and does not directly generate load prediction values, which is conducive to improving the stability of prediction results; (3) Based on the traditional similar day characteristics, add event type, direction of influence, intensity of influence and duration of effect. First, filter the same historical event days by event type to reduce the possibility of mismatch in business scenarios; (4) Using the residuals of historical similar days to form the load elasticity correction amount can correct the overestimation or underestimation of the LSTM model during business event periods; (5) Simulation results show that the overall MAE, RMSE and MAPE of the method of the present invention are 1.402MW, 1.783MW and 1.624%, respectively; on the business event day, the MAPE of the method of the present invention is 1.92%, which is 60.17% lower than the 4.82% of the traditional similar day method. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0020] Figure 1 The simulated load and meteorological variable curves provided in this embodiment of the invention; Figure 2 This is a diagram illustrating the LSTM model training process provided in an embodiment of the present invention. Figure 3 A typical daily load forecast curve provided for an embodiment of the present invention; Figure 4 This is a daily load forecast curve for business events provided in an embodiment of the present invention; Figure 5 A comparison chart of prediction errors of different methods provided in the embodiments of the present invention; Figure 6 A comparison chart of prediction errors for normal days and business event days provided in an embodiment of the present invention; Figure 7 A flowchart illustrating the load elasticity analysis method based on a large language model provided in this embodiment of the invention. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] The following detailed description of the implementation method of the present invention is in conjunction with the accompanying drawings. The description is only a partial embodiment and not all embodiments. For clarity, representations and descriptions unrelated to the present invention are omitted in the drawings and description.

[0023] To provide a clearer understanding of the technical features, objectives, and beneficial effects of this invention, the technical solutions of this invention are selected and described in detail below. Obviously, the described embodiments are only a portion of the embodiments of this invention, not all of them, and should not be construed as limiting the scope of implementation of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0024] 1. Data Construction and Preprocessing This embodiment uses MATLAB R2023b for simulation, constructing hourly data for 730 consecutive days, totaling 17,520 load points. The data includes load, temperature, humidity, weekday markers, holiday markers, and business notification text.

[0025] Business events include increased production, decreased production, equipment maintenance, temporary shutdowns, and demand response. The first 580 days are used as the training set, the middle 70 days as the validation set, and the last 80 days as the test set, arranged chronologically. Each sample inputs data from the 168 hours prior to the prediction date to predict the load for the next 24 hours.

[0026] like Figure 1 As shown, Figure 1 The model displays hourly load and meteorological variable curves for the previous 30 days. The load exhibits distinct intraday peak-valley characteristics, while temperature and humidity vary over time. Therefore, historical load, temperature, humidity, weekday indicators, and holiday indicators are used as inputs to the LSTM model.

[0027] 2. LSTM Basic Prediction The LSTM model has 5 input features, 64 hidden units, a batch size of 32, an initial learning rate of 0.001, a maximum number of training epochs of 100, and outputs the load for the next 24 hours.

[0028] like Figure 2 As shown, Figure 2 This diagram illustrates the LSTM model training process. The training and validation losses decrease rapidly in the early stages of training and then gradually stabilize, indicating that the model has essentially converged. The horizontal axis in the graph represents the number of iterations, which is not equivalent to the number of training epochs. After training, the LSTM model is used to obtain the basic prediction results for the test day.

[0029] 3. Semantic extraction of business events Input the business notifications already received in the current day into the large language model and extract the following information: Event type; Direction of influence; Influence intensity; Start time; End of the session.

[0030] The large language model only extracts event information and does not directly generate load forecasts. Event types are encoded using one-hot encoding, and the direction, intensity, and duration of influence are converted into structured features for subsequent screening of similar days.

[0031] For example, for a notification that "the production line will temporarily stop operating from 8:00 to 20:00", the large language model identifies it as a temporary production stoppage event, with a negative impact and an effective period from 8:00 to 20:00.

[0032] 4. Selection of similar dates and residual correction Traditional similar day characteristics include average temperature, average humidity, weekday sign, holiday sign, weekday attribute, and average load of the previous day.

[0033] When there are no business events on the forecast date, select similar dates from historical dates without events; when there are business events on the forecast date, first filter historical dates with the same event type, then compare the direction of influence, intensity of influence, duration of influence and traditional characteristics, and select the 5 historical dates with the highest degree of similarity.

[0034] Using the residual between the actual load on selected historical similar days and the load predicted by the LSTM, the load correction for the prediction day is calculated, and the LSTM base prediction result is adjusted accordingly. When the load on historical similar days is overestimated by the LSTM, the prediction result is adjusted downwards; when the load on historical similar days is underestimated, the prediction result is adjusted upwards.

[0035] 5. Forecast results for ordinary days and business event days like Figure 3 As shown, Figure 3 The figure shows the prediction results for an ordinary day on October 29, 2025. The figure compares the actual values, the periodic mean method, LSTM, the traditional similar day method, and the event semantic method of this invention. All methods can basically reflect the load change trend of an ordinary day. The prediction curve of the method of this invention is closer to the actual load, indicating that this method can maintain the prediction accuracy of ordinary days.

[0036] like Figure 4 As shown, Figure 4 This is the forecast result for the business event day of December 26, 2025. A temporary production stoppage occurred on this day, affecting the period from 8:00 to 20:00. Actual load decreased significantly during the event period, resulting in a substantial overestimation using the periodic mean method, and some deviations also occurred with LSTM and traditional similar day methods. This invention identifies the temporary production stoppage event and uses similar historical event days for residual correction, resulting in a forecast curve that more closely approximates the actual load.

[0037] 6. Prediction Error Analysis like Figure 5 As shown, mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are used to evaluate the prediction performance. Figure 4 The figure shows a comparison of the overall prediction errors of different methods. The results of each method are shown in Table 1.

[0038] Table 1 Comparison of overall prediction errors of different methods As shown in Table 1, the three errors of the method of the present invention are all minimized, indicating that event semantic filtering and historical residual correction can improve the overall prediction accuracy.

[0039] like Figure 6 As shown in Table 2, the MAPE comparison of different methods on normal days and business event days is presented.

[0040] Table 2 Comparison of MAPE on Normal Days and Business Event Days Using Different Methods The MAPE of the method of this invention is 1.56% on normal days and 1.92% on business event days. Compared with the traditional similar day method, the MAPE on business event days is reduced from 4.82% to 1.92%, a relative reduction of 60.17%. This shows that the present invention can effectively reduce the prediction deviation caused by business events while maintaining the prediction accuracy on normal days.

[0041] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A load elasticity analysis method based on a large language model, characterized by, Includes the following steps: S1, acquire historical load data, meteorological data, calendar attribute data and business notification text from the electricity sales company, input the historical load data, meteorological data and calendar attribute data into the long short-term memory network, and output the basic load forecast value for the next 24 hours of the forecast day; S2, input the business notification text into the large language model, and extract only the structured event semantic features of event type, direction of influence, intensity of influence, start time and end time. The large language model does not directly generate load prediction values. S3, when there is a business event on the predicted date, first filter the same type of historical event days according to the event type, then compare the direction of influence, the intensity of influence, the start time, the end time and the distance of traditional features, and select historically similar days; S4. Calculate the prediction residuals for each of the historical similar days, weight the load elasticity correction amount according to the similarity between the historical similar days and the prediction days, and add the load elasticity correction amount to the base load prediction value to obtain the final load prediction value. 2.The load elasticity analysis method based on a large language model according to claim 1, wherein, In step S2, the event type is represented by one-hot encoding, and the influence direction and influence intensity are combined into a directional influence intensity. The directional influence intensity, together with the start time and the end time, constitutes an 8-dimensional event semantic feature. When there are no business events on the prediction day, the event type encoding, influence intensity, and duration of action are all processed as if there were no events. 3.The load elasticity analysis method based on a large language model according to claim 1, characterized in that, In step S3, the traditional feature distance is calculated based on the traditional similar day feature vector. The traditional similar day feature vector includes the average temperature, average humidity, weekday indicator, holiday indicator, weekday attribute, and average load of the previous day for the predicted day, expressed as: The traditional feature distance is: ; Where, in the formula, To predict the daily average temperature, To predict the daily average humidity, As a sign of a workday As a sign for holidays, For weekdays, To predict the average load of the previous day, and These are the standardized traditional similar day features of the predicted day and the historical candidate day, respectively, where j=1,2,…,6. 4.The load elasticity analysis method based on a large language model according to claim 1, wherein, In step S3, the number of historical similar days is 5; when there is no business event on the predicted day, a historical similar day is selected from the historical dates without business events. 5.The load elasticity analysis method based on a large language model according to claim 1, wherein, In step S4, the predicted residual for time h on the k-th historical similar day is: ,in, For the actual load on historically similar days, The LSTM is used to predict the load; when the residual is greater than 0, it indicates that the Long Short-Term Memory (LSTM) network underestimates the load, and when the residual is less than 0, it indicates that the LSTM network overestimates the load. 6.The load elasticity analysis method based on a large language model according to claim 1, wherein, In step S4, the load elasticity correction for the predicted time h on day d is: ; in, For the set of historically similar days for the predicted date, For the first Normalized weights corresponding to historically similar dates For the similar day The prediction residual at each time point; The final load forecast value at time h on day d is: ,in, To predict the final load forecast value at time h on day d; This represents the base prediction value of the LSTM at the corresponding time.

7. The load elasticity analysis method based on a large language model according to claim 1, characterized in that, Before step S1, the process also includes: aligning the historical load data, meteorological data, and calendar attribute data by time, filling in missing data, and normalizing the continuous variables according to the training set parameters; using the data of the 168 consecutive hours before the prediction date as an input sample, and using the load of the next 24 hours before the prediction date as the prediction target. 8.The load elasticity analysis method based on a large language model according to claim 1, wherein, The input to the Long Short-Term Memory Network is the historical load, temperature, humidity, weekday markers, and holiday markers for the 168 consecutive hours prior to the prediction date. The input features of the Long Short-Term Memory Network are 5, the number of hidden units is 64, the batch size is 32, the initial learning rate is 0.001, and the maximum number of training rounds is 100.

9. A load elasticity analysis system based on a large language model, characterized in that, include: The data acquisition module is used to acquire historical load data, meteorological data, calendar attribute data, and business notification texts from the electricity sales company. The basic forecasting module is used to input the historical load data, meteorological data and calendar attribute data into the long short-term memory network and output the basic load forecast value for the next 24 hours of the forecast day. The event semantic extraction module is used to input the business notification text into the large language model and extract only the structured event semantic features of event type, direction of influence, intensity of influence, start time and end time. The large language model does not directly generate load prediction values. The similar day selection module is used to first filter historical event days of the same type according to the event type when there is a business event on the predicted day, and then compare the direction of influence, the intensity of influence, the start time, the end time and the distance of traditional features to select historical similar days. The correction module is used to calculate the prediction residuals for each of the historical similar days, weight the load elasticity correction amount according to the similarity between the historical similar days and the prediction days, and add the load elasticity correction amount to the base load prediction value to obtain the final load prediction value.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the load resilience analysis method based on a large language model as described in any one of claims 1 to 8.