Power load prediction method and device based on dynamic event perception, and electronic equipment

Abstract

The invention relates to a power load prediction method and device based on dynamic event awareness and electronic equipment, and the method comprises the steps: constructing a power news quantification model driven by a preset large model; based on the electric power news quantitative model, a medium and long term prediction model of time sequence decomposition and dynamic attention fusion is constructed, and the medium and long term prediction model comprises a time sequence prediction framework based on an additive model and a time sequence deep learning model based on an attention mechanism so as to capture a long-term periodicity rule and a short-term dynamic feature of the load; and outputting a predicted load result of the power system based on the dynamic sensing event by using a medium and long term prediction model of time sequence decomposition and dynamic attention fusion. Therefore, the problems that in related technologies, due to the fact that multi-modal fusion lacks an effective cross-modal space-time dynamic association mechanism and an artificial intelligence prediction model lacks interpretability analysis on key social event influence loads, deep fusion of social event semantics is difficult, response to emergencies is lagged, and interpretability is poor are solved.

Description

Technical Field

[0001] This application relates to the field of power system technology, and in particular to a power load forecasting method, apparatus and electronic equipment based on dynamic event perception. Background Technology

[0002] Medium- and long-term power load forecasting is the foundation for power grid planning and market operation decisions. Every 1% improvement in forecast accuracy can reduce operating costs by tens of millions of yuan. With China's industrial restructuring (such as the "East Data, West Calculation" strategy), the load in the western region is showing strong volatility. Traditional load forecasting models face severe challenges in dealing with the impact of complex social events such as macroeconomic regulation, extreme weather events, and changes in industrial policies.

[0003] In related technologies, load forecasting methods can be broadly categorized into three types: first, statistical models, which possess good model interpretability and low computational cost; second, probabilistic models, which can generate forecast intervals to improve forecast robustness; and third, artificial intelligence models, which have strong nonlinear feature capture capabilities. Some studies have attempted to integrate textual information reflecting social events into artificial intelligence models, and hybrid model research is gradually emerging to mitigate the error accumulation effect in multi-step forecasting.

[0004] However, among related technologies, statistical models struggle to characterize nonlinear relationships, probabilistic models are prone to biased prior assumptions, artificial intelligence models fail to effectively integrate unstructured information, text fusion techniques rely on manually constructed keyword libraries or expert priors leading to insufficient semantic depth in the representation of social events, hybrid methods lack effective cross-modal spatiotemporal dynamic correlation mechanisms resulting in weak cross-modal dynamic collaborative modeling capabilities, and artificial intelligence prediction models lack interpretable modeling mechanisms for the impact paths of social events, leading to a contradiction between predictive interpretability and engineering practicality, all of which urgently need improvement. Summary of the Invention

[0005] This application provides a power load forecasting method, device, and electronic device based on dynamic event perception to solve the problems in related technologies, such as text representation relying on manual keywords and expert experience, lack of effective cross-modal spatiotemporal dynamic correlation mechanism in multimodal fusion, lack of interpretability analysis of the impact load of key social events in artificial intelligence prediction models, resulting in difficulty in deeply integrating the semantics of social events, delayed response to emergencies, and poor interpretability.

[0006] The first aspect of this application provides a power load forecasting method based on dynamic event perception, comprising the following steps: constructing a pre-defined large-scale model-driven power news quantification model; based on the power news quantification model, constructing a medium-to-long-term forecasting model that integrates time-series decomposition and dynamic attention, wherein the medium-to-long-term forecasting model includes a time-series forecasting framework based on an additive model and a time-series deep learning model based on an attention mechanism, to capture the long-term periodicity and short-term dynamic characteristics of the load; and using the medium-to-long-term forecasting model that integrates time-series decomposition and dynamic attention to output the predicted load results of the power system based on dynamically perceived events.

[0007] Through the above-mentioned technical means, the embodiments of this application can construct a large model-driven quantitative model for power news, and combine a medium- and long-term prediction model with time-series decomposition and dynamic attention fusion to achieve power load prediction. This allows for the accurate extraction of power-related social dynamic event features from massive amounts of text, effectively realizing cross-modal fusion and interpretable architecture design of power consumption data and news text, thereby significantly improving the accuracy of medium- and long-term load prediction.

[0008] Optionally, in one embodiment of this application, the construction of a pre-defined large-scale model-driven quantitative model for electricity news includes: using the pre-defined large-scale model to analyze at least one news variable and at least one related parameter in multiple news texts that affect electricity demand; calculating secondary variable values ​​and NIQ indices based on the at least one news variable and the at least one related parameter; calculating the P-NIQ index based on the secondary variable values ​​and the NIQ index; and obtaining a quantitative representation of the impact of various social events on electricity in the news texts based on the P-NIQ index, so as to construct the quantitative model for electricity news.

[0009] Through the above-mentioned technical means, the embodiments of this application can analyze the influencing variables and parameters in news texts using a large model, calculate the values ​​of secondary variables, NIQ index and P-NIQ index, construct a quantitative model of power news, thereby deeply exploring the impact of social events on power in news texts, enhancing semantic capture capabilities, reducing subjective bias, and realizing the dynamic extraction of fine-grained event features.

[0010] Optionally, in one embodiment of this application, the formula for calculating the P-NIQ index is: , in, X i As a first-level variable, X i.j It is a second-order variable. i Number of first-level variables j Number of second-level variables n i For the first iThe number of second-level variables for each first-level variable.

[0011] Through the above-mentioned technical means, the embodiments of this application can clarify the calculation method of the P-NIQ index. By integrating multi-level variables with weights, it provides a systematic and standardized calculation basis for the quantification of the impact of social events, ensuring the scientific nature of the quantification process and the consistency of the results.

[0012] Optionally, in one embodiment of this application, the construction of a medium- to long-term prediction model based on temporal decomposition and dynamic attention fusion, wherein the medium- to long-term prediction model includes a time series prediction framework based on an additive model and a time series deep learning model based on an attention mechanism to capture the long-term periodicity and short-term dynamic characteristics of load, includes: performing correlation analysis and screening on industry-specific electricity consumption data to obtain industry electricity consumption characteristics that meet preset correlation conditions; quantifying news text data using a news quantification model to obtain news features; aligning industry electricity consumption characteristics, the news features, and economic features along the time dimension to obtain dynamic fusion features; using a Prophet model to extract annual seasonal characteristics and trend growth rates of electricity load data to obtain static features characterizing future long-term trend patterns; and inputting the static features and the dynamic fusion features into an improved TFT model for training until a preset number of training rounds are met to obtain a medium- to long-term prediction model that can be applied online to output the prediction results of future electricity load.

[0013] Through the above-mentioned technical means, the embodiments of this application can construct a medium- and long-term prediction model by screening industry electricity consumption characteristics, quantifying news characteristics, fusing multiple features and aligning them over time, combining Prophet to extract long-term static features and improving TFT model training, thereby realizing the dynamic fusion of multiple features across modes, taking into account both the long-term trend stability of load and the sensitivity to sudden social events, and improving the time-series adaptability and robustness of prediction.

[0014] Optionally, in one embodiment of this application, the evaluation formula for the medium- to long-term prediction model is: , , , , in, n To predict the total number, y i This is the actual value. These are predicted values.

[0015] Through the above-mentioned technical means, the embodiments of this application can evaluate the model based on actual values ​​and predicted values ​​using multiple evaluation formulas, comprehensively and objectively measure the accuracy of the prediction results, ensure that the prediction results have both high precision and high availability, provide a quantitative reference for model performance optimization, and guarantee the reliability and engineering applicability of the prediction results.

[0016] A second aspect of this application provides a power load forecasting device based on dynamic event perception, comprising: a quantization module for constructing a power news quantification model driven by a preset large model; a construction module for constructing a medium-to-long-term forecasting model based on the power news quantification model, wherein the medium-to-long-term forecasting model includes a time series forecasting framework based on an additive model and a time series deep learning model based on an attention mechanism; and a forecasting module for outputting the predicted load result of the power system based on dynamically perceived events using the medium-to-long-term forecasting model based on the time series decomposition and dynamic attention fusion.

[0017] Through the above-mentioned technical means, the embodiments of this application can construct a large model-driven quantitative model for power news, and combine a medium- and long-term prediction model with time-series decomposition and dynamic attention fusion to achieve power load prediction. This enables the accurate extraction of power-related social dynamic event features from massive amounts of text, effectively realizing cross-modal fusion and interpretable architecture design of power consumption data and news text, thereby significantly improving the accuracy of medium- and long-term load prediction.

[0018] Optionally, in one embodiment of this application, the quantification module includes: an analysis unit, used to analyze at least one news variable and at least one related parameter in multiple news texts that affect electricity demand using the preset large model; a calculation unit, used to calculate secondary variable values ​​and NIQ index based on the at least one news variable and the at least one related parameter; and a characterization unit, used to calculate the P-NIQ index based on the secondary variable values ​​and the NIQ index, and obtain a quantitative characterization of the impact of various social events on electricity in the news texts based on the P-NIQ index, so as to construct the electricity news quantification model.

[0019] Through the above-mentioned technical means, the embodiments of this application can analyze the influencing variables and parameters in news texts using a large model, calculate the values ​​of secondary variables, NIQ index and P-NIQ index, construct a quantitative model of power news, thereby deeply exploring the impact of social events on power in news texts, enhancing semantic capture capabilities, reducing subjective bias, and realizing the dynamic extraction of fine-grained event features.

[0020] Optionally, in one embodiment of this application, the formula for calculating the P-NIQ index is: , in, Xi As a first-level variable, X i.j It is a second-order variable. i Number of first-level variables j Number of second-level variables n i For the first i The number of second-level variables for each first-level variable.

[0021] Through the above-mentioned technical means, the embodiments of this application can clarify the calculation method of the P-NIQ index. By integrating multi-level variables with weights, it provides a systematic and standardized calculation basis for the quantification of the impact of social events, ensuring the scientific nature of the quantification process and the consistency of the results.

[0022] Optionally, in one embodiment of this application, the construction module includes: a screening unit, used to perform correlation analysis and screening on industry-specific electricity consumption data to obtain industry-specific electricity consumption characteristics that meet preset correlation conditions; a quantification unit, used to quantify news text data using a news quantification model to obtain news features; an acquisition unit, used to align industry-specific electricity consumption characteristics, the news features, and economic features along a time dimension to obtain dynamic fusion features; an extraction unit, used to extract annual seasonal features and trend growth rates of electricity load data using a Prophet model to obtain static features characterizing future long-term trend patterns; and a training unit, used to input the static features and the dynamic fusion features into an improved TFT model for training until a preset number of training rounds are met to obtain a medium- to long-term prediction model that can be applied online, so as to output the prediction results of future electricity load.

[0023] Through the above-mentioned technical means, the embodiments of this application can construct a medium- and long-term prediction model by screening industry electricity consumption characteristics, quantifying news characteristics, fusing multiple features and aligning them over time, combining Prophet to extract long-term static features and improving TFT model training, thereby realizing the dynamic fusion of multiple features across modes, taking into account both the long-term trend stability of load and the sensitivity to sudden social events, and improving the time-series adaptability and robustness of prediction.

[0024] Optionally, in one embodiment of this application, the evaluation formula for the medium- to long-term prediction model is: , , , , in, n To predict the total number, y i This is the actual value. These are predicted values.

[0025] Through the above-mentioned technical means, the embodiments of this application can evaluate the model based on actual values ​​and predicted values ​​using multiple evaluation formulas, comprehensively and objectively measure the accuracy of the prediction results, ensure that the prediction results have both high precision and high availability, provide a quantitative reference for model performance optimization, and guarantee the reliability and engineering applicability of the prediction results.

[0026] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the power load forecasting method based on dynamic event perception as described in the above embodiments.

[0027] A fourth aspect of this application provides a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described power load forecasting method based on dynamic event awareness.

[0028] A fifth aspect of this application provides a computer program product that stores a computer program that, when executed by a processor, implements the above-described power load forecasting method based on dynamic event awareness.

[0029] This application's embodiments can construct a large-scale model-driven quantitative model for power news, combined with a medium- to long-term prediction model based on temporal decomposition and dynamic attention fusion, to achieve power load forecasting. This allows for the accurate extraction of power-related social dynamic event features from massive amounts of text, effectively realizing cross-modal fusion and interpretable architecture design of power consumption data and news text, thereby significantly improving the accuracy of medium- to long-term load forecasting. This solves the problems in related technologies, such as text representation relying on manual keywords and expert experience, the lack of an effective cross-modal spatiotemporal dynamic correlation mechanism for multimodal fusion, and the lack of interpretability analysis of the impact load of key social events in artificial intelligence prediction models, leading to difficulties in deeply integrating the semantics of social events, delayed response to emergencies, and poor interpretability.

[0030] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0031] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart of a power load forecasting method based on dynamic event perception provided in an embodiment of this application; Figure 2 This is a schematic diagram of the basic framework of a TFT model provided according to an embodiment of this application; Figure 3 This is a flowchart illustrating the workflow of a medium- to long-term forecasting model provided according to an embodiment of this application. Figure 4 This is a schematic diagram of the input-output flow of a prediction model provided according to an embodiment of this application; Figure 5 A heatmap of industry attention in 2021 provided according to one embodiment of this application; Figure 6 A heatmap of news event attention in 2018 provided according to one embodiment of this application; Figure 7 A heatmap of news event attention in 2019 provided according to one embodiment of this application; Figure 8 This is a schematic diagram of the structure of a power load forecasting device based on dynamic event perception according to an embodiment of this application; Figure 9 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application.

[0032] Figure label: 10 Power load forecasting device based on dynamic event perception; 100-Quantization module, 200-Construction module, 300-Forecasting module; 901-Memory, 902-Processor, 903-Communication interface. Detailed Implementation

[0033] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0034] The following description, with reference to the accompanying drawings, describes a power load forecasting method, apparatus, and electronic device based on dynamic event perception, according to embodiments of this application. Addressing the issues mentioned in the background art, such as text representation relying on manual keywords and expert experience, lack of effective cross-modal spatiotemporal dynamic correlation mechanisms in multimodal fusion, and lack of interpretability analysis of the impact load of key social events in artificial intelligence prediction models, this application provides a power load forecasting method based on dynamic event perception. In this method, a large-model-driven power news quantification model can be constructed, combined with a medium-to-long-term prediction model based on temporal decomposition and dynamic attention fusion to achieve power load forecasting. This accurately extracts the features of power-related social dynamic events from massive amounts of text, effectively realizing cross-modal fusion and interpretable architecture design of power consumption data and news text, thereby significantly improving the accuracy of medium-to-long-term load forecasting. This addresses the problems in related technologies, such as text representation relying on manual keywords and expert experience, the lack of an effective cross-modal spatiotemporal dynamic association mechanism in multimodal fusion, and the lack of interpretability analysis of the impact load of key social events in artificial intelligence prediction models, which leads to difficulties in deeply integrating the semantics of social events, delayed response to emergencies, and poor interpretability.

[0035] Specifically, Figure 1 This is a flowchart illustrating a power load forecasting method based on dynamic event perception, provided in an embodiment of this application.

[0036] like Figure 1 As shown, the power load forecasting method based on dynamic event perception includes the following steps: In step S101, a pre-defined large-scale model-driven quantitative model for power news is constructed.

[0037] It is understood that the preset large model in the embodiments of this application can be a pre-trained large-scale language model, such as the DeepSeek large model. The preset large model can be set by those skilled in the art according to the actual situation, and no specific restrictions are made here.

[0038] In actual implementation, the embodiments of this application can acquire news text data. The acquisition sources can cover major news such as economics, international affairs, and energy, and can represent various social dynamic events. They have strong authority and representativeness. Comprehensive and representative news variables are selected to construct a power news quantitative model driven by the DeepSeek large model.

[0039] The embodiments of this application can automatically parse massive amounts of news text using large-scale pre-trained language models, achieving fine-grained and quantitative characterization of the impact of social events on power load, and providing high-quality and interpretable external event feature inputs for subsequent power load forecasting.

[0040] Optionally, in one embodiment of this application, a pre-defined large-scale model-driven power news quantification model is constructed, including: using the pre-defined large-scale model to analyze at least one news variable and at least one related parameter in multiple news texts that affect power demand; calculating secondary variable values ​​and NIQ index based on at least one news variable and at least one related parameter; calculating the P-NIQ (Power News Impact Quantifier) ​​index based on the secondary variable values ​​and NIQ index, and obtaining a quantitative representation of the impact of various social events on power in the news texts based on the P-NIQ index, so as to construct a power news quantification model.

[0041] It is understood that the news variables in the embodiments of this application may include, but are not limited to, news effectiveness, news impact, and news content; the relevant parameters may include, but are not limited to, timeliness, regional sensitivity, and environment.

[0042] For example, the specific process of constructing P-NIQ in this application embodiment can be divided into three main steps: First, analyze the news variables and related parameters that affect electricity demand in the news text through a large model; second, calculate the values ​​of secondary variables and the NIQ index; and finally, calculate the P-NIQ index through the NIQ index and the analysis results of the large model, which is a quantitative representation of the impact of various social events on electricity in the news text.

[0043] Specifically, this application embodiment uses the People's Daily to represent social dynamics. A total of 457,105 People's Daily news articles from January 2009 to December 2021 were obtained using web crawling technology. To better focus on news texts highly relevant to electricity demand, the extracted samples were first thoroughly screened. A large-scale model was used to analyze keywords in the news texts that might be related to electricity demand in western China. Then, target news articles were selected based on these keywords, ultimately identifying 5,388 valid news texts with strong relevance to electricity demand in western China.

[0044] Furthermore, in the process of selecting primary and secondary variables, this embodiment of the application can use Deepseek to conduct a comprehensive analysis of 5388 news texts: First, Deepseek is used to analyze the strength of the relationship between news characteristics and electricity, and the model's thinking process is recorded; second, the "Jieba segmentation" module in Python is used to analyze the news texts and the model's thinking process; then, high-frequency keywords are counted using ROSTCM software; finally, the keywords are analyzed and expanded using a large model to determine 5 primary variables and 21 secondary variables related to western electricity demand, as detailed in Table 1.

[0045] Table 1 Quantitative Evaluation System for Electricity News

[0046] Table 1. Quantitative Evaluation System for Electricity News (Continued)

[0047] Furthermore, based on the quantitative evaluation system for electricity news shown in Table 1, the P-NIQ index is calculated.

[0048] This application embodiment can analyze the influencing variables and parameters in news texts using a large model, calculate the values ​​of secondary variables, the NIQ index, and the P-NIQ index, and construct a quantitative model for power news. This allows for in-depth analysis of the impact of social events on power in news texts, enhancing semantic capture capabilities, reducing subjective bias, and achieving dynamic extraction of fine-grained event features.

[0049] Optionally, in one embodiment of this application, the formula for calculating the P-NIQ index is: , in, X i As a first-level variable, X i.j It is a second-order variable. i Number of first-level variables j Number of second-level variables n i For the first i The number of second-level variables for each first-level variable.

[0050] In actual implementation, the embodiments of this application can calculate the secondary variables one by one according to the evaluation criteria of each secondary variable. Then, the static news influence index NIQ and the dynamic news influence index X5 are calculated based on the values ​​of the secondary variables, as shown in the following formula: , , in, X i.j For secondary variables, NIQ is the static news influence index, and X5 is the dynamic news influence index. The P-NIQ index is calculated by combining the static and dynamic indices, as detailed in the formula below:

[0051] , in, X i As a first-level variable, X i.j It is a second-order variable. i Number of first-level variables j Number of second-level variables n i For the first iThe number of second-level variables for each first-level variable.

[0052] The embodiments of this application clearly define the calculation method of the P-NIQ index. Through multi-level variable weighted integration, it provides a systematic and standardized calculation basis for the quantification of the impact of social events, ensuring the scientific nature of the quantification process and the consistency of the results.

[0053] In step S102, based on the power news quantification model, a medium- and long-term prediction model that integrates time series decomposition and dynamic attention is constructed. The medium- and long-term prediction model includes a time series prediction framework based on an additive model and a time series deep learning model based on an attention mechanism, in order to capture the long-term periodicity and short-term dynamic characteristics of the load.

[0054] It is understood that, in the embodiments of this application, time series decomposition can be decomposing the power load time series into different features such as long-term trends and short-term fluctuations; dynamic attention fusion can be using attention mechanisms to dynamically adjust the weights of different features in the prediction, highlighting key influencing factors; the time series prediction framework based on additive models can be Prophet; and the time series deep learning model based on attention mechanisms can be TFT (Temporal Fusion Transformer, improved time series fusion transformation).

[0055] In actual implementation, the embodiments of this application can construct a medium- and long-term prediction model based on the power news quantification model, which integrates time series decomposition and dynamic attention. The medium- and long-term prediction model includes a time series prediction framework based on an additive model and a time series deep learning model based on an attention mechanism, in order to capture the long-term periodicity and short-term dynamic characteristics of load.

[0056] For example, Prophet is a time series forecasting framework based on an additive model. Its core idea is to decompose time series data into four parts: trend, seasonal, holiday, and residual components. The trend component captures the long-term evolution of electricity demand, while the seasonal component captures annual and monthly periodic patterns. The embodiments of this application can model the trend and seasonal components, which characterize the long-term evolution of load, to extract the long-term trend and periodic patterns of load data.

[0057] Considering that the relocation of heavy industry may lead to non-linear growth in the load in western China, this application adopts a piecewise linear trend model, as shown in the following equation: , in, k This represents the initial growth rate of the load. m This is the initial offset; α and β These are the increment rate adjustment vector and the offset adjustment vector, respectively.α ( t () represents the piecewise indicator function. The seasonal term is modeled using a Fourier series approximation, as shown in the following equation: , in, and Here are the Fourier coefficients, representing the amplitude and phase of the periodic component, respectively; P It is the period length; N This represents the order of the Fourier series, used to control the complexity of seasonal terms.

[0058] Through the above modeling process, the Prophet model extracts interpretable long-term trends and periodic patterns from the load sequence, providing a stable benchmark reference for the TFT model, thereby effectively mitigating the problem of error accumulation in subsequent multi-step predictions.

[0059] Furthermore, the TFT model in this embodiment can characterize static data, historical data, and future known data respectively, and its network structure is as follows: Figure 2 As shown, it consists of five main parts: gating mechanism, variable filtering network, static covariate encoder, interpretable multi-head attention and temporal fusion encoder.

[0060] Gated Residual Networks (GRNs) can adaptively uncover complex patterns, such as linear and nonlinear relationships, between data features, enhancing the model's adaptability and predictive ability under complex data conditions. It is assumed that the information it receives is the main input variable. and external static covariate vector The calculation process can be represented as follows: , , , GLU stands for gating mechanism; ω Indicates the gating weight; LayerNorm is the layer normalization operation; W i,j and b i,j These represent network weights and biases, respectively. μ 1 and μ 2 represents the intermediate layer; ELU(·) is a non-linear activation function. During training, this component retains key information through the residual network and introduces Dropout regularization to reduce overfitting, thereby improving model stability and generalization ability.

[0061] The VSN (Variable Screening Network) works similarly to an attention mechanism, ensuring the model focuses on important features through homomorphic feature weight allocation. However, the difference lies in its focus on feature selection, specifically identifying the most contributing variables from the current prediction task. For static data, historical data, and future data, this application's embodiments use VSNs with the same structure but different weights. Taking future input data as an example, assuming... Indicates the first i Variables in t The input at time t, the set of input features can be represented as: , Among them, input features E t With additional static covariates c s Weight vectors for variable selection are generated using GRN and Softmax layers. v t It can be represented as: , Each feature After entering an independent GRN for nonlinear processing, we obtain The features are then weighted and combined with the weight vector to obtain the final output features, which can be expressed as: , The static covariate encoder is responsible for transforming the long-term trend and seasonal terms output by the Prophet model into a global static representation, thereby enhancing the controllability and interpretability of time-series features. This embodiment employs a fully connected layer and LayerNorm to process the static covariate vector. P s Perform nonlinear transformations and normalization to generate low-dimensional static variables. c s , can be represented as: , in, W g This is the weight matrix. b g This is the bias term. By explicitly solidifying trends and periodic patterns, this component enables the model to inherit Prophet's interpretability decomposition capabilities in long-term predictions, while also adaptively correcting residual fluctuations through the attention mechanism of TFT, thereby mitigating the error accumulation problem of traditional hybrid models.

[0062] The MHA (Multi-Head Attention) mechanism is used to dynamically capture key features in time-series data, thereby improving model performance. Traditional TFTs capture temporal dependencies through MHA, but their attention weight calculations only rely on the autocorrelation of historical load data, without explicitly modeling the influence of external events. This application's embodiment employs the NLCMHA (News-Load Cross Multi-Head Attention) mechanism, achieving cross-modal fusion by dynamically aligning news event embeddings with load temporal features. Firstly, dynamic news features... E news Compared with the original load time series characteristics E load By mapping the vectors to the same feature space through linear transformations, attention query vectors are generated. Q Key vector K Sum value vector V They are represented as follows: , , , in W Q , W K , W V This is a learnable weight matrix. Then, the attention weights are calculated using a scaled dot product attention mechanism. A The expression is as follows: , in, d h The dimension of the attention head is used to scale the dot product result to stabilize gradient computation. Through the above steps, the model can dynamically capture the cross-modal correlation between news event features and load time-series features, thereby quantifying the impact of social events (such as power rationing, epidemics, natural disasters, etc.) on load forecasting. After calculating the attention weights, the value vector is processed through the MHA mechanism. V Weighted summaries are applied to generate cross-modal fusion feature representations. The NLCMHA mechanism employs a multi-head structure to capture feature correlations in different subspaces. Each attention head independently computes attention weights and generates corresponding outputs. The multi-head results are then fused through additive aggregation to generate the final features. , can be represented as: , in hThe number of attention heads is [not specified]. NLCMHA enables the model to capture multi-scale associations between news events and load features in different subspaces, thereby improving its ability to model the impact of complex social events. Furthermore, to ensure the model meets the causal constraints of time series prediction, this embodiment introduces a causal mask into the NLCMHA mechanism. By constructing an upper triangular mask matrix, the attention mechanism is restricted to focusing only on information from the current and historical time steps, while masking the impact of future time steps. [Mask matrix] M i,j The definition is as follows: , This mask matrix is ​​used in the calculation of the attention score to ensure that the news attention weight is zero at future time steps, thus avoiding information leakage. When predicting future loads, this causal constraint ensures that the model relies only on historical data and the news features of the current input, meeting the needs of real-world prediction scenarios.

[0063] The temporal fusion decoder consists of a two-layer GRN network and an improved attention mechanism. First, the outputs of static covariates (trend / seasonal features extracted by Prophet) and dynamic variables are enhanced by the first-layer GRN to represent temporal features. Second, NLCMHA quantifies the impact of news events through a dynamic weight matrix and applies a decoder mask to ensure that each time step focuses only on historical events. Then, the output of the attention layer is residually connected to the enhanced features and input into the second-layer GRN for nonlinear compression. Finally, an adaptive skip connection is introduced, allowing the output of the sequence-to-sequence layer to pass directly to the final prediction layer, improving adaptability to high-fluctuation scenarios.

[0064] The embodiments of this application can achieve collaborative modeling of long-term patterns and short-term fluctuations by fusing temporal decomposition and dynamic attention, thereby improving the model's adaptability to complex load changes and, in particular, enhancing its sensitivity to sudden social events.

[0065] Optionally, in one embodiment of this application, a medium- to long-term prediction model based on time-series decomposition and dynamic attention fusion is constructed. The medium- to long-term prediction model includes a time-series prediction framework based on an additive model and a time-series deep learning model based on an attention mechanism to capture the long-term periodicity and short-term dynamic characteristics of load. This includes: performing correlation analysis and screening on industry-specific electricity consumption data to obtain industry electricity consumption characteristics that meet preset correlation conditions; quantifying news text data using a news quantification model to obtain news features; aligning industry electricity consumption characteristics, news features, and economic features along the time dimension to obtain dynamic fusion features; using a Prophet model to extract annual seasonal characteristics and trend growth rates of electricity load data to obtain static features that characterize future long-term trend patterns; and inputting the static features and dynamic fusion features into an improved TFT model for training until a preset number of training rounds are met to obtain a medium- to long-term prediction model that can be applied online, thereby outputting the prediction results of future electricity load.

[0066] It is understood that the preset correlation condition in the embodiments of this application can be, and the preset training rounds can be, respectively. For example, embodiments of this application can construct a medium- to long-term prediction model that integrates time-series decomposition and dynamic attention to capture the long-term periodicity and short-term dynamic characteristics of the load. The workflow diagram of the medium- to long-term prediction model is shown below. Figure 4 As shown, it consists of four parts: multi-feature fusion, static feature extraction, model training, and load prediction.

[0067] Specifically, in this application embodiment, the industry electricity consumption data can first be screened by correlation analysis to obtain industry electricity consumption characteristics with strong correlation; then, news text data can be quantified by news quantification model to obtain news characteristics; finally, industry electricity consumption characteristics, news characteristics and economic characteristics can be aligned with time dimension to obtain dynamic fusion characteristics.

[0068] Furthermore, embodiments of this application can construct a medium- to long-term prediction model by screening industry electricity consumption characteristics, quantifying news characteristics, fusing multiple features and aligning them over time, combining Prophet to extract long-term static features and improving TFT model training, thereby achieving multi-feature cross-modal dynamic fusion, taking into account both the long-term trend stability of load and the sensitivity to sudden social events, and improving the time-series adaptability and robustness of prediction.

[0069] This application embodiment can construct a medium-to-long-term prediction model by screening industry electricity consumption characteristics, quantifying news characteristics, fusing multiple features and aligning them over time, combining Prophet to extract long-term static features and improving TFT model training, thereby achieving multi-feature cross-modal dynamic fusion, taking into account both the long-term trend stability of load and the sensitivity to sudden social events, and improving the time-series adaptability and robustness of prediction.

[0070] In step S103, the predicted load results of the power system are output based on the dynamic sensing events using a medium-to-long-term prediction model that combines time-series decomposition and dynamic attention fusion.

[0071] It is understood that, in the embodiments of this application, the dynamically perceived events can be information for the model to identify and respond to external social events (such as policy promulgation, extreme weather, industrial relocation, etc.) in real time.

[0072] For example, in this embodiment of the application, industry electricity consumption data, news data, and historical economic data can be used as known data; future known data can be monthly data; and static covariate data can be long-term trend terms and seasonal terms. The model uses these three types of data to predict electricity demand for the next several months. The prediction time scale of the model based on dynamically perceived events can be adaptively adjusted according to demand. The input and output process of the prediction model is as follows: Figure 3 As shown, the input consists of three parts: historical known data, future known data, and static covariate data. When the trained medium- to long-term forecasting model is fed real-time fused data, the model can quickly output the forecast results of future electricity load, achieving accurate prediction of electricity consumption over multiple months.

[0073] The embodiments of this application can achieve high-precision and interpretable online forecasting of medium- and long-term loads, providing real-time and reliable decision-making basis for power grid dispatching and planning.

[0074] Optionally, in one embodiment of this application, the evaluation formula for the medium- to long-term forecasting model is: , , , , in, n To predict the total number, y i This is the actual value. These are predicted values.

[0075] In practical implementation, the embodiments of this application can predict performance using the evaluation formula of the medium- and long-term prediction model, selecting MAE (Mean Absolute Error), MAPE (Mean Absolute Percentage Error), RMSE (Root Mean Square Error), and NRMSE (Normalized Root Mean Square Error) for evaluation. The lower the evaluation value, the smaller the error. The specific calculation formula is as follows: , , , , in, n To predict the total number, y i This is the actual value. These are predicted values.

[0076] The embodiments of this application can evaluate the model based on actual and predicted values ​​using multiple evaluation formulas, comprehensively and objectively measuring the accuracy of the prediction results, ensuring that the prediction results have both high precision and high availability, providing a quantitative reference for model performance optimization, and guaranteeing the reliability and engineering applicability of the prediction results.

[0077] Specifically, it can be combined with Figures 4 to 7 As shown, the working principle of the power load forecasting method based on dynamic event perception in this application is explained in detail with a specific embodiment.

[0078] This application uses actual monthly load data, economic data (including GDP and PPI), and news data from a province in western China from 2009 to 2021 as the original dataset to verify the effectiveness of the proposed load forecasting method. The data from 2009 to 2020 is used as the training set to train the model and optimize the parameters, and the data from 2020 to 2021 is used as the test set to evaluate the model's forecasting performance.

[0079] To verify the superiority of the proposed TDDA-MLLF (Temporal Decomposition and Dynamic Attention for Medium- and Long-term Load Forecasting) model in capturing the dynamic characteristics of monthly load, a single-month load forecasting experiment was conducted in this embodiment. The experiment compared six time-series models that integrate quantitative features of electricity news, including N-RNN, N-LSTM, N-Transformer, N-Prophet, N-TFT, and the TDDA-MLLF proposed in this embodiment. All models used a 12-month sliding window to forecast the electricity load for 2021 month by month. The evaluation results are shown in Table 2.

[0080] Table 2. Performance Comparison Results of Six Models in Monthly Load Forecasting

[0081] As shown in Table 2, the proposed TDDA-MLLF model performs best in the monthly forecasting task. Compared to the second-best performing N-TFT model, TDDA-MLLF reduces MAE by approximately 34.9% and MAPE by approximately 33.0%. TDDA-MLLF also demonstrates a significant advantage in RMSE and NRMSE metrics, with RMSE and NRMSE reduced by approximately 40.7% compared to N-TFT, indicating that TDDA-MLLF not only has high prediction accuracy but also a more concentrated distribution of prediction errors and better stability, validating the effectiveness of the time-series decomposition module Prophet in improving load forecasting. Furthermore, the deep learning models with attention mechanisms (TDDA-MLLF, N-TFT, and N-Transformer) outperform the traditional statistical model N-Prophet and the conventional time-series deep learning models N-LSTM and N-RNN, highlighting the effectiveness of attention mechanisms in handling time-series dependencies.

[0082] To verify the error suppression capability of the TDDA-MLLF model over a long time scale, this application presents an embodiment of a multi-month load forecasting experiment. The experiment uses a sliding window strategy to directly predict the load of the above six models for the next 12 months (single output of the annual forecast results), and the evaluation results are shown in Table 3.

[0083] Table 3. Performance Comparison Results of Six Models in Multi-Month Load Forecasting

[0084] As shown in Table 3, TDDA-MLLF still performs best in multi-month forecasting tasks, achieving the best results across all metrics. Specifically, the prediction error metrics MAE and MAPE are reduced by 16.1% and 17.3%, respectively, compared to the second-best model, N-Prophet. Compared to single-month forecasting, the N-Prophet model exhibits a smaller prediction error than N-TFT in multi-month forecasting, demonstrating the advantage of the time-series decomposition module in multi-step forecasting. Furthermore, traditional time-series deep learning models N-LSTM and N-RNN still perform the worst across all forecasting metrics, indicating that conventional time-series deep learning models are insufficient to effectively mitigate the long-term error accumulation problem in load forecasting.

[0085] To verify the effective impact of electricity news on load forecasting, an ablation experiment of news embedding was conducted. Four sets of comparative experiments were constructed by controlling the news input type: full-news, retaining only positive news, retaining only negative news, and no news input. The comparative results generated by the multi-month forecasting model are shown in Table 4.

[0086] Table 4 Performance Comparison Results of Four News Input Types

[0087] As shown in Table 4, the type of news input has a significant impact on load prediction. The model performs best in the scenario with complete news input, which indicates that the P-NIQ model can effectively quantify news and thus achieve cross-modal fusion of social event information. In the scenario without news embedding, the model performance drops significantly, with the prediction error increasing by 78.4% compared to the scenario with complete news, further illustrating the necessity of considering social event features in load prediction. In the experiment with incomplete news embedding, although both positive and negative news can improve the effect of load prediction, the improvement effect of positive news is more significant. This is because there are more positive news items in the news sample, which can provide the model with more comprehensive dynamic information.

[0088] This application's embodiments aim to study the impact of the TDDA-MLLF model on the attention levels of different industries. An experiment was conducted to investigate the influence of attention visualization technology. The experiment generated a 2021 industry attention heatmap based on news-load cross-attention weights. For example... Figure 5 As shown, the model focuses most on the chemical raw material manufacturing industry (CE), which aligns closely with the province's positioning as a national energy and chemical industry base. Looking at the attention levels across different months, this industry received relatively high attention in March, September, and October 2021. This is attributed to the tightening of energy consumption control policies in March and the production line shutdowns caused by environmental inspections in September and October, triggering a surge in attention weights. This demonstrates the model's adaptive ability to capture sudden load changes in this industry. Furthermore, the model pays the least attention to other industries with the lowest correlation to load trends, effectively filtering out fluctuations in non-dominant industry data.

[0089] This application embodiment studies the impact of news events on the model and generates a cross-year news event heatmap through attention weights. Figure 6 and Figure 7 The images show heatmaps of news events for 2018 and 2019, with the horizontal axis representing the month the news occurred (the current year) and the vertical axis representing the predicted target month (the following year). Figure 6 It can be seen that news events in September 2018 had the greatest impact on the load forecast for 2019. This is because the province released a policy in September to deepen the reform of "streamlining administration, delegating power, and improving services," which stimulated market vitality and led to an increase in high-energy-consuming enterprises. The impact of January and February 2018 on the load forecast for 2019 was second. Although there were no major news releases in January and February 2018, we observed that the quantitative values ​​of news events in January and February from 2015 to 2017 were significantly lower than those in March, while the quantitative values ​​of news events in January and February 2018 were on par with March. This indicates that the model captured the historical pattern of influence between news months in 2018, thus assigning higher weights to January and February. Figure 7It can be seen that news events in November and December 2019 had a significant impact on the load forecast for 2020. In November 2019, the province's reservoirs transported water to support winter irrigation in the Yellow River Basin, causing a surge in electricity consumption in the Yellow River irrigation area of ​​the province in 2020. December 2019 was the early stage of the COVID-19 outbreak, and there was still relatively little news related to the epidemic. However, the model captured the impact of the epidemic on the electricity load of various industries in 2020, and therefore gave December a higher weight.

[0090] The embodiments of this application can focus on important emergencies in the current month and better capture the impact of emergencies on subsequent loads, thereby providing more comprehensive information for load forecasting under extreme events.

[0091] The power load forecasting method based on dynamic event perception proposed in this application can construct a large-scale model-driven power news quantification model. Combined with a medium- to long-term forecasting model integrating temporal decomposition and dynamic attention fusion, it achieves power load forecasting, thereby accurately extracting power-related social dynamic event features from massive amounts of text. This effectively realizes cross-modal fusion and interpretable architecture design of power consumption data and news text, significantly improving the accuracy of medium- to long-term load forecasting. This solves the problems in related technologies, such as text representation relying on manual keywords and expert experience, lack of effective cross-modal spatiotemporal dynamic correlation mechanisms in multimodal fusion, and lack of interpretability analysis of the impact load of key social events in artificial intelligence prediction models. These issues lead to difficulties in deeply integrating the semantics of social events, delayed responses to emergencies, and poor interpretability.

[0092] Next, with reference to the accompanying drawings, a power load forecasting device based on dynamic event perception proposed according to an embodiment of this application is described.

[0093] Figure 8 This is a schematic diagram of the power load forecasting device based on dynamic event perception according to an embodiment of this application.

[0094] like Figure 8 As shown, the power load forecasting device 10 based on dynamic event perception includes: a quantization module 100, a construction module 200, and a forecasting module 300.

[0095] Among them, the quantification module 100 is used to construct a power news quantification model driven by a pre-set large model; Module 200 is used to build a medium- and long-term prediction model based on the power news quantification model, which integrates time series decomposition and dynamic attention. The medium- and long-term prediction model includes a time series prediction framework based on an additive model and a time series deep learning model based on an attention mechanism. The prediction module 300 is used to output the predicted load results of the power system based on the dynamic sensing events using a medium- and long-term prediction model that combines time-series decomposition and dynamic attention fusion.

[0096] Optionally, in one embodiment of this application, the quantization module 100 includes: an analysis unit, a calculation unit, and a characterization unit.

[0097] The analysis unit is used to analyze at least one news variable and at least one related parameter in multiple news texts that affect electricity demand using a pre-set large model.

[0098] The calculation unit is used to calculate the values ​​of secondary variables and the NIQ index based on at least one news variable and at least one relevant parameter.

[0099] The representation unit is used to calculate the P-NIQ index based on the values ​​of secondary variables and the NIQ index, and to obtain a quantitative representation of the impact of various social events on electricity in the news text based on the P-NIQ index, so as to construct a quantitative model of electricity news.

[0100] Optionally, in one embodiment of this application, the formula for calculating the P-NIQ index is: , in, X i As a first-level variable, X i.j It is a second-order variable. i Number of first-level variables j Number of second-level variables n i For the first i The number of second-level variables for each first-level variable.

[0101] Optionally, in one embodiment of this application, the construction module 300 includes: a screening unit, a quantization unit, an extraction unit, and a training unit.

[0102] The filtering unit is used to perform correlation analysis and filtering on electricity consumption data by industry to obtain industry electricity consumption characteristics that meet preset correlation conditions.

[0103] The quantization unit is used to quantify news text data using a news quantization model to obtain news features.

[0104] The acquisition unit is used to align industry electricity consumption characteristics, news characteristics, and economic characteristics along the time dimension to obtain dynamically fused characteristics.

[0105] The extraction unit is used to extract the annual seasonal characteristics and trend growth rate of electricity load data using the Prophet model, so as to obtain static features that characterize the long-term trend.

[0106] The training unit is used to input static features and dynamic fusion features into the improved TFT model for training until the preset training rounds are met, so as to obtain a medium- and long-term prediction model that can be applied online and output the prediction results of future power load.

[0107] Optionally, in one embodiment of this application, the evaluation formula for the medium- to long-term forecasting model is: , , , , in, n To predict the total number, y i This is the actual value. These are predicted values.

[0108] It should be noted that the foregoing explanation of the embodiment of the power load forecasting method based on dynamic event perception also applies to the power load forecasting device based on dynamic event perception in this embodiment, and will not be repeated here.

[0109] The power load forecasting device based on dynamic event perception proposed in this application can construct a large-scale model-driven quantitative model for power news, and combine it with a medium- to long-term forecasting model based on temporal decomposition and dynamic attention fusion to achieve power load forecasting. This allows for the accurate extraction of power-related social dynamic event features from massive amounts of text, effectively realizing cross-modal fusion and interpretable architecture design of power consumption data and news text, thereby significantly improving the accuracy of medium- to long-term load forecasting. This solves the problems in related technologies, such as text representation relying on manual keywords and expert experience, the lack of an effective cross-modal spatiotemporal dynamic correlation mechanism for multimodal fusion, and the lack of interpretability analysis of the impact load of key social events in artificial intelligence prediction models. These issues lead to difficulties in deeply integrating the semantics of social events, delayed response to emergencies, and poor interpretability.

[0110] Figure 9 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 901, the processor 902, and the computer program stored on the memory 901 and capable of running on the processor 902.

[0111] When the processor 1202 executes the program, it implements the power load forecasting method based on dynamic event perception provided in the above embodiments.

[0112] Furthermore, electronic devices also include: Communication interface 1203 is used for communication between memory 1201 and processor 1202.

[0113] The memory 1201 is used to store computer programs that can run on the processor 1202.

[0114] The memory 1201 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage.

[0115] If the memory 901, processor 902, and communication interface 903 are implemented independently, then the communication interface 903, memory 901, and processor 902 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0116] Optionally, in a specific implementation, if the memory 1201, processor 1202, and communication interface 1203 are integrated on a single chip, then the memory 1201, processor 1202, and communication interface 1203 can communicate with each other through an internal interface.

[0117] The processor 1202 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0118] This embodiment also provides a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described power load forecasting method based on dynamic event awareness.

[0119] This application also provides a computer program product storing a computer program that, when executed by a processor, implements the above-described power load forecasting method based on dynamic event awareness.

[0120] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0121] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0122] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0123] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0124] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0125] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0126] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0127] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A power load forecasting method based on dynamic event perception, characterized in that, Includes the following steps: Construct a quantitative model for power news driven by a pre-defined large model; Based on the aforementioned power news quantification model, a medium- to long-term prediction model integrating time series decomposition and dynamic attention is constructed. The medium- to long-term prediction model includes a time series prediction framework based on an additive model and a time series deep learning model based on an attention mechanism, in order to capture the long-term periodicity and short-term dynamic characteristics of load. The medium- to long-term prediction model, which combines time-series decomposition and dynamic attention fusion, outputs predicted load results for the power system based on dynamically perceived events.

2. The method according to claim 1, characterized in that, The construction of the pre-defined large-scale model-driven quantitative model for power news includes: The preset large model is used to analyze at least one news variable and at least one related parameter that affect electricity demand in multiple news texts. Calculate the secondary variable value and NIQ index based on the at least one news variable and the at least one related parameter; The P-NIQ index is calculated based on the values ​​of the secondary variables and the NIQ index, and a quantitative representation of the impact of various social events on electricity in news texts is obtained based on the P-NIQ index, so as to construct the quantitative model of electricity news.

3. The method according to claim 2, characterized in that, The formula for calculating the P-NIQ index is as follows: ， in, X i As a first-level variable, X i.j It is a second-order variable. i Number of first-level variables j Number of second-level variables n i For the first i The number of second-level variables for each first-level variable.

4. The method according to claim 1, characterized in that, The construction of a medium- to long-term prediction model that integrates temporal decomposition and dynamic attention, comprising a time-series prediction framework based on an additive model and a time-series deep learning model based on an attention mechanism, aims to capture the long-term periodicity and short-term dynamic characteristics of the load, including: Correlation analysis and screening of electricity consumption data by industry segment are conducted to obtain industry electricity consumption characteristics that meet preset correlation conditions; News text data is quantified using a news quantification model to obtain news features; Align industry electricity consumption characteristics, news characteristics, and economic characteristics along the time dimension to obtain dynamically fused characteristics; The Prophet model was used to extract the annual seasonality and trend growth rate of electricity load data to obtain static features that characterize the long-term trend. The static features and the dynamic fusion features are input into the improved TFT model for training until a preset number of training rounds are met, resulting in a medium- to long-term prediction model that can be applied online, so as to output the prediction results of future power load.

5. The method according to claim 1 or 4, characterized in that, The evaluation formula for the medium- and long-term forecasting model is as follows: ， ， ， ， in, n To predict the total number, y i This is the actual value. These are predicted values.

6. A power load forecasting device based on dynamic event perception, characterized in that, include: The quantification module is used to build a pre-defined large-scale model-driven quantitative model for power news. The module is used to construct a medium- and long-term prediction model based on the power news quantification model, which integrates time series decomposition and dynamic attention. The medium- and long-term prediction model includes a time series prediction framework based on an additive model and a time series deep learning model based on an attention mechanism. The prediction module is used to output the predicted load results of the power system based on the dynamic sensing events using the medium- and long-term prediction model that combines the time-series decomposition and dynamic attention fusion.

7. The apparatus according to claim 6, characterized in that, The quantization module includes: The analysis unit is used to analyze at least one news variable and at least one related parameter in multiple news texts that affect electricity demand using the preset large model; A calculation unit is used to calculate the secondary variable value and the NIQ index based on the at least one news variable and the at least one related parameter; The characterization unit is used to calculate the P-NIQ index based on the values ​​of the secondary variables and the NIQ index, and to obtain a quantitative characterization of the impact of various social events on electricity in the news text based on the P-NIQ index, so as to construct the electricity news quantitative model.

8. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the power load forecasting method based on dynamic event perception as described in any one of claims 1-5.

9. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the power load forecasting method based on dynamic event awareness as described in any one of claims 1-5.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed to implement the power load forecasting method based on dynamic event awareness as described in any one of claims 1-5.

Images