CN-122000888-A - Electric load prediction method, device, equipment and medium based on hybrid expert framework

CN122000888ACN 122000888 ACN122000888 ACN 122000888ACN-122000888-A

Abstract

The invention discloses an electric load prediction method based on a hybrid expert framework, which relates to the technical field of electric load prediction and is used for solving the problem of inaccurate existing prediction; the method comprises the steps of pre-training a general electric load prediction expert model, storing model parameters obtained after training as initialization parameters, fine tuning based on the initialization parameters, generating route embedding through class perception contrast learning, dynamically distributing input electric load samples to corresponding specialized expert models according to the route embedding, and predicting an electric load time sequence through the distributed specialized electric load prediction expert model. The invention also discloses an electric load prediction device, electronic equipment and a computer storage medium based on the hybrid expert framework. According to the invention, by introducing a specialized expert framework and a class perception contrast routing mechanism, the prediction accuracy of the electric load is further improved.

Inventors

WANG HUIDONG
CHEN LINHONG
WU HAOTIAN
CHENG YUAN
WU HAO
RUAN JIAN
WANG JUN
YIN HONGYUAN
XIAO JIDONG
FANG ZHICHUN
LI LEI
LU PENGFEI
CAO RUIFENG
ZHANG YUNLEI
JIANG CHI
YE HONGDOU

Assignees

国网浙江省电力有限公司营销服务中心

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. An electric load prediction method based on a hybrid expert framework is characterized by comprising the following steps: Acquiring an electric load time sequence training data set, and constructing a prediction framework composed of a plurality of electric load prediction expert models, wherein the electric load prediction expert models at least comprise a general electric load prediction expert model and a plurality of specialized electric load prediction expert models; Pre-training a general electric load prediction expert model, inputting the electric load time sequence training data set into the general electric load prediction expert model for training, and storing model parameters obtained after training is completed as initialization parameters; Based on the initialization parameters, the biased specialized electrical load prediction expert model is finely adjusted by combining a low-rank adaptation mechanism; Generating a routing embedding with electric load mode discrimination capability through class perception contrast learning, and dynamically distributing an input electric load sample to a corresponding partial specific electric load prediction expert model according to the routing embedding; and predicting the electric load time sequence through the distributed metaspecialization electric load prediction expert model.
2. The method for predicting electric load based on mixed expert architecture as claimed in claim 1, wherein the electric load time series sample is input into a general electric load prediction expert model in an end-to-end manner for training, so that the general electric load prediction expert model learns the electric load commonality dynamic characteristics of cross-region and cross-time scale, and after training, model parameters are taken as the initialization parameters.
3. The hybrid expert architecture based electrical load prediction method of claim 1, wherein the initialization parameters include a weight parameter and a bias parameter of a multi-headed self-attention layer, and a predicted head weight parameter and a bias parameter for electrical load numerical output.
4. A hybrid expert architecture based electrical load prediction method as claimed in claim 1 or 3, wherein fine tuning the meta-specific electrical load prediction expert model in combination with a low rank adaptation mechanism based on the initialization parameters comprises: And respectively performing specialized fine tuning on the plurality of specialized electric load prediction expert models by utilizing parameters of the general electric load prediction expert model, so as to satisfy the following conditions: , wherein, Parameters representing a generic electrical load prediction expert model, Represents the first Parameters of individual specialized electrical load prediction experts; In the fine tuning process, a low-rank adaptation mode is adopted, only part of parameter subspaces are updated to maintain general knowledge and enhance the adaptation capability to specific time sequence modes, and in the fine tuning process, the first is The output calculations of the individual electrical load prediction experts satisfy: , wherein, A time series is represented by a sequence of time, Representing a low rank delta matrix, Representing the passing of After fine tuning, the first The output result finally obtained by the individual electric load prediction expert; The low rank delta matrix Implemented by the product of two learnable matrices, expressed as: Wherein , , To be smaller than the original parameter dimension Is an inherent rank of (c).
5. The hybrid expert architecture-based electrical load prediction method of claim 1, wherein generating a routing embedding with electrical load pattern discrimination capability through class-aware contrast learning comprises: extracting a potential representation of the input electrical load time series samples; defining a contrast loss function and mapping the potential representation as a decision-embedding through a routing network; obtaining a routing embedding with the electrical load mode discrimination capability by utilizing the representation capability of contrast learning optimization decision embedding so as to guide sample distribution; Wherein the contrast loss function Is calculated as follows: , Wherein, the Is a similarity function, τ is a temperature hyper-parameter, The point of the anchor is indicated and, A positive sample is represented and a positive sample is represented, Representing a negative sample.
6. The method for predicting electrical loads based on mixed expert architecture according to claim 1, wherein predicting electrical load time series through allocated specialized electrical load prediction expert models comprises calculating probability distribution of electrical load time series samples to be predicted, which are allocated to each specialized electrical load prediction expert model through Softmax function according to corresponding route embedding, allocating the samples to one or more specialized electrical load prediction expert models according to the probability distribution to predict, and weighting and summarizing a plurality of prediction results to obtain final electrical load prediction results.
7. The hybrid expert architecture-based electrical load prediction method of claim 1, wherein the generic electrical load prediction expert model and the meta-specific electrical load prediction expert model are both built based on PatchTST time-series modeling architecture.
8. An electrical load prediction device based on a hybrid expert architecture, characterized in that it comprises: The construction module is used for acquiring an electric load time sequence training data set and constructing a prediction framework composed of a plurality of electric load prediction expert models, wherein the electric load prediction expert models at least comprise a general electric load prediction expert model and a plurality of bias specific electric load prediction expert models; The training module is used for pre-training the universal electric load prediction expert model, inputting the electric load time sequence training data set into the universal electric load prediction expert model for training, and storing model parameters obtained after training is completed as initialization parameters; generating a route embedding with the electric load mode distinguishing capability through class perception contrast learning, and dynamically distributing an input electric load sample to a corresponding specialized electric load prediction expert model according to the route embedding; and the prediction module is used for predicting the electrical load time sequence through the distributed metaspecialization electrical load prediction expert model.
9. An electronic device comprising a processor, a storage medium and a computer program stored in the storage medium, characterized in that the computer program, when executed by the processor, implements the hybrid expert architecture based electrical load prediction method of any one of claims 1 to 7.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the hybrid expert architecture based electrical load prediction method of any one of claims 1 to 7.

Description

Electric load prediction method, device, equipment and medium based on hybrid expert framework Technical Field The invention relates to the technical field of electric load prediction, in particular to an electric load prediction method, device, equipment and medium based on a hybrid expert framework for electric load prediction and a specialized expert framework. Background The electric load prediction is a key basic technology for the operation and the scheduling of an electric power system, and the core aim is to accurately grasp the dynamic characteristics and the internal rules of the evolution of the electric load along with time through deep mining and analysis of historical electricity consumption data, so that the effective prediction of the electric power demand in a certain period in the future is realized. The technology plays a vital role in guaranteeing the safety and stability of the power grid, improving the energy economic efficiency and promoting the clean energy fusion, and is particularly embodied in a plurality of practical application scenes such as power grid planning, power generation resource optimal scheduling, demand side response management, high-proportion new energy consumption and the like. However, electrical load time series often exhibit a high degree of complexity and uncertainty in the actual environment. The change is not only affected by basic time sequence trend and cycle rule, but also by coupling action of external multiple factors, including meteorological conditions (such as temperature and humidity), seasonal replacement, user electricity behavior mode, time-of-use electricity price policy, holidays, emergencies and the like. The influence of these factors makes the load sequence generally exhibit strong non-stationarity, multi-scale fluctuation and obvious mode heterogeneity. In other words, there are often large differences in statistical distribution, variation cadence and fluctuation patterns of electrical load curves corresponding to different regions, different industry users, different electrical properties (such as industry, business, residents) and even different time periods (such as weekdays and weekends, peaks and valleys), which brings great challenges to the construction of a high-precision prediction model with wide adaptability. At present, the prior art has made a certain progress in the aspects of prediction model architecture and feature engineering, but the existing method mainly relies on a globally unified and parameter-sharing modeling mode, namely training and deducing load sequences of all types and scenes by using the same group of model parameters. When facing load modes with significant differences in practice, the method is easy to cause mode confusion and mutual interference of the model in the learning process, and ideal prediction precision is difficult to maintain in various different scenes, so that generalization performance and robustness of the model in a complex real power environment are restricted. Therefore, a new time series prediction method capable of accurately identifying and characterizing the intrinsic heterogeneity of an electrical load sequence and realizing differential modeling and adaptive inference for different load modes is needed. The overall accuracy, stability and practical value of the electric load prediction technology in diversified and diversified actual application scenes are improved. Disclosure of Invention In order to overcome the defects of the prior art, one of the purposes of the invention is to provide an electric load prediction method based on a hybrid expert framework, which is characterized in that an electric load general prediction expert model is firstly trained to learn commonalities of electric load changes in different areas and scenes through a two-stage hybrid expert model framework oriented to electric load prediction, and then an input electric load time series sample is intelligently distributed to a plurality of electric load specialized prediction expert models subjected to high-efficiency fine tuning by combining a class sensing contrast routing mechanism, so that fine modeling and prediction are realized for different electric load modes. One of the purposes of the invention is realized by adopting the following technical scheme: an electric load prediction method based on a hybrid expert framework comprises the following steps: Acquiring an electric load time sequence training data set, and constructing a prediction framework composed of a plurality of electric load prediction expert models, wherein the electric load prediction expert models at least comprise a general electric load prediction expert model and a plurality of specialized electric load prediction expert models; Pre-training a general electric load prediction expert model, inputting the electric load time sequence training data set into the general electric load prediction expert model for training, so that th