Short-term load multi-scenario prediction method based on large language model event driving

Abstract

This invention relates to the field of power system load forecasting and discloses a short-term load multi-scenario forecasting method based on a large language model and event-driven approach. The method includes acquiring historical load data and decomposing it into trend, periodic, and random components; collecting and preprocessing multi-source text information related to the forecast period; outputting event type labels, time impact intensity levels, and forecast uncertainty level parameters based on a pre-trained large language model; using a conditional generative adversarial network (CGAN) with conditional vectors and random noise vectors as inputs to perform parallel forecasting of the trend, periodic, and random components, and performing multiple sampling and reconstructions based on Monte Carlo simulations to generate a set of forecast scenarios containing multiple load curves; performing statistical analysis on the generated set of forecast scenarios to output load forecasting results with probabilistic information. The advantage of this invention is its insensitivity to sample size, making it suitable for load forecasting in small-sample scenarios.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and in particular to a short-term load multi-scenario forecasting method based on event-driven large language models. Background Technology

[0002] Load forecasting plays a crucial role in maintaining power system stability. Accurate forecasting helps maintain the supply-demand balance of the power system. With the development of artificial intelligence, load forecasting has achieved high accuracy through machine learning and deep learning methods. However, due to the special nature and confidentiality of load data, it is often difficult to obtain sufficient load data, leading to insufficient training of forecasting models and poor forecast accuracy. As the penetration rate of renewable energy continues to increase, the operating environment of the power system exhibits significant uncertainties, and unstructured text information such as weather changes, holiday arrangements, and grid operation announcements has an increasingly significant impact on load fluctuations.

[0003] For short-term load forecasting, there are generally three approaches: statistical and time series methods. Statistical methods construct explicit mathematical models by analyzing the time dependence, trend, and seasonality of load series, such as ARIMA and Holt-Winters. These methods typically assume that the load series has a certain degree of stationarity and periodicity, achieving short-term forecasting through linear modeling. These models have advantages such as clear structure, strong interpretability, and low computational cost, but they struggle to handle strongly nonlinear behavior and have limited ability to model exogenous variables. Traditional machine learning methods learn the mapping relationship between load and various influencing factors through manual feature engineering and data-driven approaches. Typical models include support vector regression, random forests, and gradient boosting trees. These methods are suitable for multidimensional input data, possessing strong nonlinear fitting capabilities and high prediction accuracy. However, they rely heavily on feature engineering and struggle to automatically capture complex time series dependency structures. Deep learning methods utilize neural networks to automatically extract features and model nonlinear relationships in time series, representing the mainstream direction of load forecasting research in recent years. These include sequence models based on LSTM / GRU, convolutional models based on CNN / TCN, and attention mechanism models based on Transformer. Deep learning possesses powerful expressive capabilities, simultaneously utilizing historical load, meteorological, and contextual information. However, these methods typically require large amounts of data for training and have relatively weak interpretability.

[0004] Existing short-term load forecasting methods mainly rely on historical load and structured meteorological data, making it difficult to effectively utilize textual information. Some methods introduce large language models to directly participate in load forecasting calculations, but they generally suffer from the following problems: large language models directly output forecast results, lacking interpretability and stability; textual information is only used as an additional feature input, failing to characterize the differentiated impact of different events on load forecasting strategies; and most methods only output a single forecast value, failing to reflect forecast uncertainty and hindering scheduling decisions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a short-term load multi-scenario prediction method based on event-driven large language model.

[0006] The objective of this invention is achieved through the following technical solution: a short-term load multi-scenario prediction method based on a large language model event-driven approach, the method comprising,

[0007] S1. Obtain historical load data and perform signal decomposition on the historical load data to obtain trend components, periodic components and random components;

[0008] S2. Collect and preprocess multi-source text information related to the forecast period, wherein the multi-source text information includes at least weather forecast text, holiday arrangement text, and power grid operation announcement text;

[0009] S3. Based on a pre-trained large language model, the processed text information is used for event recognition and uncertainty assessment, and the event type label, time impact intensity level and prediction uncertainty level parameters are output.

[0010] S4. Encode the event type label and event impact intensity level into a condition vector; at the same time, according to the prediction uncertainty level parameter, adaptively adjust the variance or sampling range of the random noise vector used for scenario generation, and determine the number of multiple scenarios generated.

[0011] S5. Using the Conditional Generative Adversarial Network (CGAN), the conditional vector and random noise vector are used as inputs to perform parallel predictions on the trend component, periodic component and random component. Based on Monte Carlo simulation, multiple sampling and reconstruction are performed to generate a set of prediction scenarios containing multiple load curves.

[0012] S6. Perform statistical analysis on the generated prediction scenario set and output load prediction results with probability information.

[0013] Specifically, the variational mode decomposition algorithm is used to decompose the historical load data into signals. The variational mode decomposition algorithm is expressed as follows:

[0014] ;

[0015] In the formula, and These are the modal components obtained from the decomposition and their corresponding center frequencies; * represents the Dirac function; * represents the convolution operation; The square of the L2 norm; The number of modes obtained by decomposing the signal; For time variables The first derivative; The imaginary unit; The first decomposition obtained One modal function;

[0016] A secondary penalty factor was introduced. and Lagrange multipliers The original problem is transformed into an unconstrained augmented Lagrangian function; the alternating direction multiplier method (ADMM) is used for iterative optimization in the frequency domain.

[0017] ;

[0018] ;

[0019] ;

[0020] In the formula, This represents the original signal in the frequency domain. For the first Frequency domain of each modal component; This represents the Lagrange multipliers in the frequency domain. It is the angular frequency variable; For the first In the nth iteration, the 1st Frequency domain solution for each mode; For the first In the nth iteration, the 1st The central angular frequency of each mode; For the first In each iteration, the frequency domain update result of the Lagrange multipliers;

[0021] Based on the center frequency of each mode The modes are sorted and frequency thresholds are set. Modes below the threshold are classified as trend components, modes within the periodic frequency band are classified as periodic components, and other high-frequency modes are regarded as random fluctuation components.

[0022] Specifically, the preprocessing of multi-source text information in S2 includes standardizing and vectorizing the multi-source text information to form text embedding vectors.

[0023] Specifically, during the training process of the large language model, the pre-training parameters and the underlying information encoding module of the large language model are frozen, and only the top-level task adaptation layer is fine-tuned.

[0024] Specifically, in S5, independent generator networks are established for the trend component, periodic component, and random component, respectively. Each generator network process follows a minimax game mechanism, and its objective function is:

[0025] ;

[0026] In the formula, D is the discriminator; G is the generator; This is a sample of the actual load components; This is a conditional vector of event type and intensity extracted from a large language model; This refers to random noise that follows a specific distribution. For real samples Calculate the mathematical expectation; To handle random noise and conditions Perform joint sampling;

[0027] Each generator network outputs trend predictions, periodic predictions, and random predictions, respectively.

[0028] Specifically, in S5, based on Monte Carlo simulation, multiple sampling and reconstruction are performed to generate a set of prediction scenarios containing multiple load curves. This includes: independently sampling the random noise space multiple times according to a pre-set number of multi-scenario generation and a noise variance parameter controlled by the uncertainty level; for each sampled noise vector, each generator network calculates a corresponding prediction subsequence through forward propagation; subsequently, a linear superposition reconstruction operation is performed, accumulating the trend prediction value, periodic prediction value, and random prediction value under the same sampling index point by point, as shown below:

[0029] ;

[0030] In the formula, This is a trend forecast value; This is a periodic forecast value; These are random predicted values;

[0031] go through The final output of the next reconstruction iteration includes A set of predicted scenarios for possible load curves.

[0032] Specifically, step S6 involves processing the data generated in S5. Statistical analysis of the time-series cross-section of the possible load curves is performed to calculate the probability quantiles at each time point, and a prediction interval with a clear confidence level is constructed.

[0033] The present invention has the following advantages:

[0034] This invention achieves event-driven, uncertainty-aware short-term load forecasting in small-sample scenarios by integrating the event recognition capabilities of a large language model with the generative capabilities of a conditional generative adversarial network. This method not only effectively utilizes unstructured text information to improve prediction accuracy and interpretability, but also adaptively generates multi-scenario probabilistic predictions, significantly enhancing the power system's decision support capabilities in dealing with uncertainty. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the prediction method of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0037] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0038] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0039] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0040] like Figure 1As shown, a short-term load multi-scenario prediction method based on event-driven large language models is proposed. This method includes:

[0041] S1. Obtain historical load data and perform signal decomposition on the historical load data to obtain trend components, periodic components and random components;

[0042] Variational mode decomposition (VMD) algorithm is used to decompose historical load data into signals. VMD is a non-recursive, adaptive signal processing technique that decomposes complex raw signals by solving variational problems. Decomposed into Each mode has a sparse intrinsic mode function (IMF). They are all tightly clustered around a central frequency. The algorithm minimizes the sum of the estimated bandwidths of each mode while ensuring that the sum of the bandwidths of all modes equals the original input signal, thus achieving effective separation and noise reduction of the original signal in the frequency domain. To obtain these bandwidth-constrained sub-signals, the variational mode decomposition algorithm first uses the Hilbert transform to calculate the analytic signal of each mode to obtain the one-sided spectrum, and then mixes the center frequency term... The spectra of each mode are shifted to baseband. Based on this, a constrained variational problem is constructed, aiming to minimize the sum of the estimated bandwidths of each mode, while constraining the sum of all modes to equal the original input signal. This optimization objective function can be mathematically expressed as:

[0043] ;

[0044] In the formula, and These are the modal components obtained from the decomposition and their corresponding center frequencies; * represents the Dirac function; * represents the convolution operation; The square of the L2 norm; The number of modes obtained by decomposing the signal; For time variables The first derivative; The imaginary unit; The first decomposition obtained One modal function;

[0045] To solve the aforementioned optimization problem with constraints, a quadratic penalty factor is introduced. and Lagrange multipliers The original problem is transformed into an unconstrained augmented Lagrangian function; the alternating direction multiplier method (ADMM) is used for iterative optimization in the frequency domain.

[0046] ;

[0047] ;

[0048] ;

[0049] In the formula, This represents the original signal in the frequency domain. For the first Frequency domain of each modal component; This represents the Lagrange multipliers in the frequency domain. It is the angular frequency variable; For the first In the nth iteration, the 1st Frequency domain solution for each mode; For the first In the nth iteration, the 1st The central angular frequency of each mode; For the first In each iteration, the frequency domain update result of the Lagrange multipliers;

[0050] Based on the center frequency of each mode The modes are sorted and frequency thresholds are set. Modes below the threshold are classified as trend components, modes within the periodic frequency band are classified as periodic components, and the remaining modes are regarded as random fluctuation components.

[0051] The above process not only effectively separates the signal frequencies but also eliminates high-frequency noise through Wiener filtering. Once the algorithm converges, based on the spectral characteristics and physical meaning of each decomposed mode, it is reorganized into three categories of load components with clear interpretability. Specifically, low-frequency modes are aggregated into trend components, reflecting long-term load growth and seasonal baselines; mid-frequency modes constitute periodic components, characterizing daily or weekly cyclical fluctuations caused by human production and lifestyle patterns; and modes containing high-frequency information are classified as random components, mainly covering uncertainties caused by meteorological changes or system noise. Through this hierarchical decoupling, the originally complex mixed load sequence is transformed into distinctive subsequences, providing high-quality input features for subsequent multi-scenario prediction.

[0052] S2. Collect and preprocess multi-source text information related to the forecast period. The multi-source text information includes at least weather forecast text, holiday arrangement text, and power grid operation announcement text. The preprocessing includes cleaning, time alignment, semantic parsing, and structured processing of the collected text information.

[0053] The key to multimodal load forecasting lies in the effective extraction and utilization of latent features from unstructured data. Therefore, this invention aims to establish a systematic mechanism for the collection and standardized processing of multi-source heterogeneous information. This mechanism first acquires comprehensive textual data on external environmental factors within the forecast period, covering key dimensions such as weather forecasts, holiday arrangements, and power grid operation announcements. Unlike traditional numerical inputs, meteorological texts not only contain numerical records of temperature and humidity but also include semantic descriptions of sudden weather changes (such as cold waves and typhoon warnings). Holiday texts record detailed information on statutory holidays, work schedule adjustments, and large-scale social events. Power grid announcement texts focus on operational instructions such as planned maintenance, line faults, and temporary power rationing. This broad data collection strategy aims to capture sudden or socially impactful factors that are difficult to reflect in pure time-series data, thereby providing rich information context for subsequent event-driven analysis. Preprocessing of multi-source textual information includes standardization and vectorization to form text embedding vectors.

[0054] S3. Based on a pre-trained large language model, the processed text information is used for event recognition and uncertainty assessment, and the event type label, time impact intensity level and prediction uncertainty level parameters are output.

[0055] After standardizing and preprocessing multi-source information, a deep logical mapping between unstructured text descriptions and power grid operating states is established using the powerful semantic reasoning capabilities of Large Language Models (LLM). Unlike traditional methods that simply concatenate text as auxiliary feature vectors into the time-series model, this invention proposes a decoupled "interpretation-decision" architecture. Specifically, a second-generation generative pre-trained model (GPT-2) is selected as the backbone network for semantic understanding, with text embedding vectors generated in S2 serving as input. To overcome the problem of scarce training data, especially in small-sample scenarios related to power load data, a transfer learning strategy is employed during training: most of the pre-training parameters and underlying information encoding modules of GPT-2 are frozen, with only the top-level task adaptation layer being fine-tuned. This strategy not only maximizes the preservation of the model's general language knowledge and logical reasoning capabilities learned on large-scale corpora but also effectively prevents overfitting on specific small-sample tasks.

[0056] The large language model does not directly perform regression tasks to predict specific load values. Instead, it acts as a high-level decision support unit, responsible for extracting key event features and uncertainty indicators from complex environmental descriptions. By combining load trend features extracted from historical time-series data, the large language model performs deep attention calculations and contextual analysis on the input text sequence, ultimately outputting decision parameters in three dimensions: First, an event type label, where the model automatically identifies the specific scenario category corresponding to the current text description (such as "extreme high-temperature day," "major holiday work schedule adjustment," or "sudden power grid maintenance"); second, an event impact intensity level, quantifying the potential impact of the event on the shape of the power grid load curve; and finally, a prediction uncertainty level parameter, which reflects the model's confidence assessment of the load fluctuation range under the current environment.

[0057] S4. Encode the event type label and event impact intensity level into a condition vector; at the same time, according to the prediction uncertainty level parameter, adaptively adjust the variance or sampling range of the random noise vector used for scenario generation, and determine the number of multiple scenarios generated.

[0058] After receiving the event semantic labels and uncertainty assessment parameters output by the large language model, this step acts as a crucial bridge connecting qualitative analysis and quantitative prediction. It aims to achieve precise "one-scenario-one-policy" predictions through an adaptive policy scheduling mechanism. This mechanism dynamically adjusts two external parameters based on the prediction uncertainty level parameters output by the LLM: the variance (or sampling range) of the random noise z and the number N of multiple scenarios generated, thus controlling the divergence and coverage of the prediction results. To effectively capture the complex distribution patterns of load data in small-sample scenarios and generate diverse prediction scenarios, this invention introduces a Conditional Generative Adversarial Network (CGAN) as the core prediction engine. CGAN employs a generator... With discriminator The game theory framework encodes the event type labels and their impact intensity levels identified by the large language model into condition vectors. This mechanism is deeply embedded in the generator's input layer; it forces the generator to strictly follow the semantic constraints of external events while capturing data distribution, thereby ensuring that the generated load curve is highly consistent with the current holiday, weather, or power grid operation status in terms of shape.

[0059] To address the randomness control and multi-scenario generation of prediction results, the system constructs an adaptive sampling mechanism based on the uncertainty level parameter. The CGAN generation process relies on a random noise vector following a specific distribution (such as a Gaussian distribution). To introduce diversity, i.e., to generate results The uncertainty level of the LLM output directly controls the variance or sampling range of the random noise parameter: when the uncertainty of the assessment is high (such as in the event of sudden extreme weather), the algorithm will expand the noise vector. The search space is expanded, and the number of multi-scenario generation is significantly increased, prompting the model to explore potential extreme load scenarios, thereby constructing a more comprehensive probability prediction interval. Conversely, on stable days with higher certainty, the noise variance is reduced and the number of samplings is decreased to obtain more concentrated prediction results. In this way, the model can not only generate multiple statistically reasonable load curves for each event label, but also dynamically adjust the divergence of the prediction results, providing panoramic decision support for power grid dispatching that includes confidence information.

[0060] S5. Using a Conditional Generative Adversarial Network (CGAN), the conditional vector and random noise vector are used as inputs to perform parallel predictions on the trend component, periodic component, and random component. Based on Monte Carlo simulation, multiple sampling and reconstruction are performed to generate a set of prediction scenarios containing multiple load curves. Based on the three orthogonal modal components of trend component, periodic component, and random component analyzed in S1, this invention adopts a divide-and-conquer strategy to establish independent generator networks for each of the trend component, periodic component, and random component, and to perform parallel predictions on each component.

[0061] To enable the generative model to accurately learn the distribution characteristics of load components under different environmental backgrounds, the training process of each sub-model follows a minimax game mechanism, and its objective function is defined as follows:

[0062] ;

[0063] In the formula, D is the discriminator; G is the generator; This is a sample of the actual load components; This is a conditional vector of event type and intensity extracted from a large language model; This refers to random noise that follows a specific distribution. For real samples Calculate the mathematical expectation; To handle random noise and conditions Joint sampling is performed; within this game theory framework, the discriminator... Attempting to use known external event conditions Under the constraints, maximize the differentiation of the actual load components. Generator-generated fake samples The ability; and the generator The goal is to minimize the probability of the discriminator detecting the error. Through this adversarial training, the temporal feature vector and event condition vector generated in the preceding steps are deeply fused, forcing the generator's predictions to conform to both historical physical evolution and explicit constraints from the current external environment (such as sudden weather changes or holiday effects). Particularly for the random component, the generator utilizes controlled random noise introduced into the formula. It can effectively capture the uncertainty characteristics in high-frequency fluctuations, thereby simulating a variety of potential random disturbance patterns; and outputs trend prediction values, periodic prediction values ​​and random prediction values ​​through each generator network.

[0064] To construct statistically significant multi-scenario prediction results, multiple sampling and reconstruction are performed based on Monte Carlo simulation to generate a prediction scenario set containing multiple load curves. This involves multiple independent samplings of the random noise space based on a pre-set number of multi-scenario generators and a noise variance parameter adjusted by the uncertainty level. For each sampled noise vector, each generator network calculates a corresponding prediction subsequence through forward propagation. Subsequently, a linear superposition reconstruction operation is performed, accumulating the trend prediction value, periodic prediction value, and random prediction value under the same sampling index point by point, as shown below:

[0065] ;

[0066] In the formula, This is a trend forecast value; This is a periodic forecast value; These are random predicted values;

[0067] go through The final output of the next reconstruction iteration includes The set of predicted scenarios for possible load curves is provided. This result set not only provides the expected value of point predictions, but also constructs probability intervals at different confidence levels through scenario distribution. It provides grid dispatchers with panoramic decision support covering the best, worst and most likely situations, and significantly enhances the system's robustness in dealing with unknown risks under small sample sizes.

[0068] S6. Perform statistical analysis on the generated prediction scenario set and output load prediction results with probability information.

[0069] After completing component reconstruction and generating a large-scale load forecast scenario set, the core task of this stage is to transform these discrete stochastic processes into statistical indicators and decision-making bases that can be directly analyzed by the power grid dispatching system. Unlike traditional single numerical forecasts that only provide deterministic point estimates, this system outputs a forecast result set containing rich probabilistic information. This is achieved through the generation of S5... A hypothetical load curve is subjected to statistical analysis across its time-series cross-section to calculate the probability quantiles at each moment, thereby constructing a prediction interval with a clear confidence level (e.g., 90% or 95%). The median is typically considered the baseline load curve for deterministic scheduling, while the upper and lower quantiles depict the boundary ranges of potential grid load fluctuations driven by current weather and social events. This representation intuitively quantifies the prediction risk introduced by the sparsity of small sample data and the uncertainty of the external environment.

[0070] The proposed method was compared with existing mainstream time series forecasting methods, with the training data duration gradually increasing from 1 month to 7 months. Table 1 shows the accuracy performance on the test set. As can be seen from Table 1, when the training data is 1 month long, the proposed method's accuracy on the test set is at least 1 percentage point higher than the other three comparison methods. While the accuracy of the other three comparison methods generally increases with the increase in training data, the accuracy of the proposed method does not change significantly. This further demonstrates that the proposed method is insensitive to sample size and is suitable for load forecasting in small sample scenarios.

[0071] Table 1 Comparison of performance of different models under different sample sizes

[0072]

[0073] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.

Claims

1. A short-term load multi-scenario prediction method based on event-driven large language model, characterized by: The method includes, S1. Obtain historical load data and perform signal decomposition on the historical load data to obtain trend components, periodic components and random components; S2. Collect and preprocess multi-source text information related to the forecast period, wherein the multi-source text information includes at least weather forecast text, holiday arrangement text, and power grid operation announcement text; S3. Based on a pre-trained large language model, the processed text information is used for event recognition and uncertainty assessment, and the event type label, time impact intensity level and prediction uncertainty level parameters are output. S4. Encode the event type label and event impact intensity level into a condition vector; at the same time, according to the prediction uncertainty level parameter, adaptively adjust the variance or sampling range of the random noise vector used for scenario generation, and determine the number of multiple scenarios generated. S5. Using the Conditional Generative Adversarial Network (CGAN), the conditional vector and random noise vector are used as inputs to perform parallel predictions on the trend component, periodic component and random component. Based on Monte Carlo simulation, multiple sampling and reconstruction are performed to generate a set of prediction scenarios containing multiple load curves. S6. Perform statistical analysis on the generated prediction scenario set and output load prediction results with probability information.

2. The method of claim 1, wherein the method is characterized by: The variational mode decomposition algorithm is used to decompose the historical load data into signals. The variational mode decomposition algorithm is expressed as follows: ； In the formula, and These are the modal components obtained from the decomposition and their corresponding center frequencies; * represents the Dirac function; * represents the convolution operation; The square of the L2 norm; The number of modes obtained by decomposing the signal; Represents the time variable The first derivative; The imaginary unit; The first decomposition obtained One modal function; A secondary penalty factor was introduced. and Lagrange multipliers The original problem is transformed into an unconstrained augmented Lagrangian function; the alternating direction multiplier method (ADMM) is used for iterative optimization in the frequency domain. ； ； ； In the formula, This represents the original signal in the frequency domain. For the first Frequency domain of each modal component; This represents the Lagrange multipliers in the frequency domain. It is the angular frequency variable; For the first In the nth iteration, the 1st Frequency domain solution for each mode; For the first In the nth iteration, the 1st The central angular frequency of each mode; For the first In each iteration, the frequency domain update result of the Lagrange multipliers; According to the center frequency of each mode The modes are ranked and a frequency threshold is set, modes below the threshold are classified as trend components, modes within the periodic band are considered as periodic components, and the rest are considered as random fluctuation components.

3. The method of claim 1, wherein the method is characterized by: The preprocessing of multi-source text information in S2 includes standardizing and vectorizing the multi-source text information to form text embedding vectors.

4. The short-term load multi-scenario prediction method based on large language model event-driven method according to claim 1, characterized in that: During the training process, the pre-training parameters and underlying information encoding modules of the large language model are frozen, and only the top-level task adaptation layer is fine-tuned.

5. The method of claim 1, wherein the method is characterized by: In S5, independent generator networks are established for the trend component, periodic component, and random component, respectively. Each generator network follows a minimax game mechanism, and its objective function is: ； In the formula, D is the discriminator; G is the generator; This is a sample of the actual load components; This is a conditional vector of event type and intensity extracted from a large language model; This refers to random noise that follows a specific distribution. For real samples Calculate the mathematical expectation; To handle random noise and conditions Perform joint sampling; Each generator network outputs trend predictions, periodic predictions, and random predictions, respectively.

6. The method of claim 5, wherein the method further comprises: In S5, multiple sampling and reconstruction based on Monte Carlo simulation are performed to generate a prediction scenario set containing multiple load curves. This includes multiple independent samplings in the random noise space based on a pre-set number of multi-scenario generation and a noise variance parameter controlled by the uncertainty level. For each sampled noise vector, each generator network calculates a corresponding prediction subsequence through forward propagation. Subsequently, a linear superposition reconstruction operation is performed, accumulating the trend prediction value, periodic prediction value, and random prediction value under the same sampling index point by point, as shown below: ； In the formula, is a trend forecast value; is a cycle forecast value; is a random forecast value; After the last reconstruction iteration, the final output contains a set of predicted scenarios of the load curve.

7. The method of claim 6, wherein the method further comprises: The specific steps of S6 are as follows: [The steps described in S6 are missing from the original text and can be omitted.] Statistical analysis of the time-series cross-section of the possible load curves is performed to calculate the probability quantiles at each time point, and a prediction interval with a clear confidence level is constructed.

Images