CN-121638591-B - Power load and electricity price combined prediction method and system based on multitask learning

CN121638591BCN 121638591 BCN121638591 BCN 121638591BCN-121638591-B

Abstract

The invention relates to the technical field of joint prediction, in particular to a method and a system for joint prediction of power load and electricity price based on multi-task learning. The method comprises the steps of respectively executing time sequence decomposition on a historical power load sequence and a historical power price sequence, constructing a multi-scale joint input feature vector sequence, obtaining shared feature representation, generating task biased shared features by a learnable channel attention module, obtaining preliminary power load features and preliminary power price features, generating deeply fused load task features and power price task features by a cross-task interaction gating unit, generating power load and power price predicted values, calculating load predicted loss, power price predicted loss and relation consistency loss, obtaining a joint total loss function, and updating network parameters. The scheme of the invention can relieve task conflict and negative migration, excavate deep relation between power load and electricity price, ensure consistency of prediction results and real conditions and reduce difficulty of network learning.

Inventors

CAO PENG
ZHANG JIGUANG
WANG JIAYU
CAO XUEZHOU
LIU YILONG
SUN HUI
LIN SEN

Assignees

陕西省水电开发集团股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260130

Claims (8)

1. The power load and electricity price combined prediction method based on multitask learning is characterized by comprising the following steps of: The method comprises the steps of obtaining a historical power load sequence, a historical power price sequence, meteorological data and calendar data, respectively executing time sequence decomposition on the historical power load sequence and the historical power price sequence to obtain respective trend components, periodic components and residual components, splicing the decomposed components with the meteorological data and the calendar data, and constructing a multi-scale joint input feature vector sequence; The method comprises the steps of inputting a multi-scale joint input feature vector sequence into a shared feature extraction network to obtain a shared feature representation, utilizing a learnable channel attention module to adjust the channel weight of the shared feature representation, and generating a task biased shared feature; The method comprises the steps of respectively inputting shared characteristics of task deflection into a power load exclusive network and a power price exclusive network to obtain preliminary power load characteristics and preliminary power price characteristics, utilizing a cross-task interaction gating unit to fuse the preliminary power load characteristics and the preliminary power price characteristics to generate deeply fused load task characteristics and power price task characteristics, generating load gating signals by the aid of a third full-connection layer and a Sigmoid activation function, performing element grading on the load gating signals and the preliminary power load characteristics to obtain load characteristics modulated by power price information, generating power price gating signals by the aid of a fourth full-connection layer and a Sigmoid activation function, performing element grading on the preliminary power price gating signals and the preliminary power price characteristics to obtain power price characteristics modulated by the load information, adding the preliminary power load characteristics and the load characteristics modulated by the power price information to generate deeply fused load task characteristics, adding the preliminary power price characteristics and the power price characteristics modulated by the load information to generate deeply fused power price task characteristics; based on the deeply fused load task feature and electricity price task feature, generating a power load predicted value and an electricity price predicted value at a target moment through a prediction head network; The method for calculating the relational consistency loss comprises the steps of calculating the pearson correlation coefficient between a predicted power load sequence and a predicted power price sequence, calculating the pearson correlation coefficient between a real power load sequence and a real power price sequence, and taking the absolute value of the difference between the two pearson correlation coefficients as the relational consistency loss.
2. The method for jointly predicting power load and power rate based on multi-task learning according to claim 1, wherein the performing time-series decomposition on the historical power load sequence and the historical power rate sequence to obtain the trend component, the periodic component and the residual component respectively comprises: The method comprises the steps of adopting a decomposition algorithm based on a local weighted regression scattered point smoothing method, decomposing a historical power load sequence input into a load trend component, a load period component and a load residual component, and decomposing a historical power price sequence input into a power price trend component, a power price period component and a power price residual component.
3. The method for jointly predicting power load and electricity price based on multi-task learning according to claim 1, wherein the step of inputting the multi-scale joint input feature vector sequence into a shared feature extraction network to obtain a shared feature representation comprises the following steps: The shared feature extraction network adopts a bidirectional gating circulation unit network, the multi-scale combined input feature vector sequence is input into the bidirectional gating circulation unit network, time sequence features are respectively extracted from the front direction and the back direction through the forward and the backward gating circulation unit layers, and the states of the forward hidden layers and the backward hidden layers are spliced to form a shared feature representation.
4. The method for joint prediction of power load and electricity price based on multi-task learning according to claim 1, wherein the adjusting the channel weight of the shared feature representation by the learning channel attention module to generate the task biased shared feature comprises: The method comprises the steps of carrying out extrusion operation on the shared feature representation through a global average pooling layer to obtain a channel indicator, inputting the channel indicator into a bottleneck structure formed by two full-connection layers to carry out excitation operation, learning weight coefficients of all channels, multiplying the learned weight coefficients with the shared feature representation according to the channels, and generating task biased shared features.
5. The method for jointly predicting power load and power price based on multi-task learning according to claim 1, wherein the step of inputting the shared characteristic of task bias to the power load task exclusive network and the power price task exclusive network to obtain a preliminary power load characteristic and a preliminary power price characteristic comprises the steps of: And respectively inputting the shared characteristics of the task bias to a first full-connection layer serving as an exclusive network of the power load task and a second full-connection layer serving as an exclusive network of the electricity price task to obtain the primary power load characteristics and the primary electricity price characteristics.
6. The method for jointly predicting power load and power price based on multi-task learning according to claim 5, wherein generating the predicted value of power load and predicted value of power price at the target moment through the prediction head network based on the deeply fused load task feature and power price task feature comprises: The prediction head network is a first multi-layer perceptron and a second multi-layer perceptron which are the same in structure, and the multi-layer perceptron is composed of two full-connection layers and a linear output layer; And the deep-fused electricity price task features are input to a second multi-layer perceptron to output the electricity price predicted value at the target moment.
7. The method for combined prediction of power load and power rate based on multitasking learning of claim 1, wherein said load prediction loss and power rate prediction loss are both calculated using an average absolute error function.
8. The utility model provides a power load and price of electricity joint prediction system based on multitask study which characterized in that includes: A processor and a memory storing computer program instructions for joint prediction of power load and power rate based on multi-tasking, which when executed by the processor implement the joint prediction method of power load and power rate based on multi-tasking according to any of claims 1-7.

Description

Power load and electricity price combined prediction method and system based on multitask learning Technical Field The invention relates to the technical field of joint prediction. More particularly, the invention relates to a method and a system for jointly predicting power load and electricity price based on multitasking learning. Background The power load can reflect the power consumption requirement of a user, the power price is influenced by various factors such as supply and demand relation, power generation cost, market strategy and the like, and the power load and the power price show a mutual influence relation on time sequence. In the case of studying the power load and electricity price prediction problem, researchers typically consider the power load and electricity price prediction as two independent single-task prediction problems, i.e., respectively constructing prediction models for the two tasks. However, the above-described treatment method breaks the inherent connection between the power load and the electricity price, and fails to fully utilize the interaction mechanism of the power load and the electricity price in time sequence. In order to take advantage of the correlation between power load and electricity price, the prior art employs a multitasking learning framework for joint prediction of power load and electricity price. In a specific network structure, a bottom network structure with shared hard parameters is generally adopted, namely, two tasks share a feature extractor and are connected with independent prediction heads, and the method tries to realize information sharing among different tasks to a certain extent by sharing the feature extraction network of the bottom layer, so that the accuracy of prediction is improved. However, the method of the multitasking learning framework is capable of linking the power load with the electricity price through the sharing of information, but has obvious disadvantages. Firstly, the hard parameter sharing structure forces all tasks to use the same characteristic representation, so that preference differences of different tasks on shared characteristics are difficult to distinguish, gradient conflict and negative migration phenomena are easy to occur among the tasks, and the training effect of the joint prediction model is influenced. Secondly, the existing joint prediction model generally builds joint loss by adding the loss functions of tasks to be simple, and the mode lacks modeling and constraint on a correlation structure between power load and electricity price prediction sequences, so that the prediction result cannot be guaranteed to be consistent with the real situation at the relation level. In addition, the original time sequence containing multiple complex modes such as trend, periodicity, random fluctuation and the like is directly used as the input of the joint prediction model, so that the difficulty of network learning and distinguishing different scale features is increased, and the extraction capability of the joint prediction model on the complex time sequence features is limited. Disclosure of Invention The invention aims to provide a power load and electricity price combined prediction method and system based on multi-task learning, which are used for solving the problems that gradient conflict and negative migration phenomenon are easy to occur between different tasks in the prior art, the prediction result cannot be guaranteed to be consistent with the real situation in a relation level, and the built combined prediction model has weak extraction capacity on complex features. In a first aspect, the present invention provides a method for jointly predicting power load and electricity price based on multitasking learning, including: The method comprises the steps of obtaining a historical power load sequence, a historical power price sequence, meteorological data and calendar data, respectively executing time sequence decomposition on the historical power load sequence and the historical power price sequence to obtain respective trend components, periodic components and residual components, splicing the decomposed components with the meteorological data and the calendar data to form a multi-scale combined input feature vector sequence, inputting the multi-scale combined input feature vector sequence into a shared feature extraction network to obtain a shared feature representation, adjusting channel weights of the shared feature representation by a learnable channel attention module to generate task biased shared features, respectively inputting the task biased shared features into a power load task exclusive network and a power price task exclusive network to obtain preliminary power load features and preliminary power price features, utilizing a cross-task interaction gating unit to fuse the preliminary power load features and the preliminary power price features to generate deep fused load task features and power pr