CN-122022879-A - New energy station electricity price prediction optimization method and system based on multisource fusion learning

CN122022879ACN 122022879 ACN122022879 ACN 122022879ACN-122022879-A

Abstract

The invention provides a new energy station electricity price prediction optimization method and system based on multi-source fusion learning, which relate to the technical field of electricity price prediction, and the method comprises the steps of constructing a dynamic space-time diagram topology based on heterogeneous electricity price driving factor sets; after the LSTM unit is distributed, node state iterative update is carried out through a hierarchical collaborative learning mechanism, a power price prediction graph neural network model is output, heterogeneous modal feature fusion coding is carried out, K+Q dynamic feature coding vectors are obtained, topology collaborative evolution reasoning of the power price prediction graph is carried out, a power price optimization prediction result comprising a power price prediction value and an uncertainty quantization interval is output, and the power price optimization prediction result is sent to a new energy station transaction decision end to refer to generation of a power market bidding strategy and risk control. The method solves the technical problem that the prediction result accuracy is insufficient due to the fact that the prior art generally adopts a single data source prediction model and the correlation among various driving factors cannot be effectively captured in the complex and changeable power market environment.

Inventors

AI YANG
YIN YING
Yao Wancan
NIE HUI
HUO XIAO
Tan Zhuohang
LU ZHIFAN
XIE HONGHUI

Assignees

江西联合能源有限公司

Dates

Publication Date: 20260512
Application Date: 20251219

Claims (10)

1. The new energy station electricity price prediction optimization method based on multi-source fusion learning is characterized by comprising the following steps of: constructing a dynamic space-time diagram topology based on the heterogeneous electricity price driving factor set; After distributing LSTM units to K core prediction nodes and Q auxiliary feature nodes of the dynamic space-time diagram topology, carrying out node state iterative updating of the dynamic space-time diagram topology through a hierarchical collaborative learning mechanism, and outputting an electricity price prediction diagram neural network model; heterogeneous mode feature fusion coding is carried out on the K+Q real-time driving factor data according to the K+Q attribute embedded characterizations, and K+Q dynamic feature coding vectors are obtained; Loading the K+Q dynamic feature coding vectors to K core prediction nodes and Q auxiliary feature nodes of the electricity price prediction graph neural network model, executing topology collaborative evolution reasoning of the electricity price prediction graph, and outputting an electricity price optimization prediction result, wherein the electricity price optimization prediction result comprises an electricity price prediction value and an uncertainty quantization interval; And sending the electricity price optimization prediction result to a new energy station transaction decision-making end, and referring to generation of an electricity market bidding strategy and risk control.
2. The new energy station electricity price prediction optimization method based on multi-source fusion learning according to claim 1, wherein a dynamic space-time diagram topology is constructed based on a heterogeneous electricity price driving factor set, and the method comprises: Carrying out power price fluctuation association capturing on new energy power price multielement influencing factors to obtain the heterogeneous power price driving factor set, wherein the heterogeneous power price driving factor set comprises K core driving factors and Q auxiliary driving factors; And after setting K core prediction nodes and Q auxiliary characteristic nodes based on the K core driving factors and the Q auxiliary driving factors respectively, carrying out node-by-node dynamic binding based on dynamic edge relation rules to complete the construction of the dynamic space-time diagram topology.
3. The new energy station electricity price prediction optimization method based on multi-source fusion learning according to claim 2, wherein after distributing LSTM units to K core prediction nodes and Q auxiliary feature nodes of the dynamic space-time diagram topology, node state iterative updating of the dynamic space-time diagram topology is performed through a hierarchical collaborative learning mechanism, and an electricity price prediction diagram neural network model is output, the method comprises: after distributing LSTM units to K core prediction nodes and Q auxiliary feature nodes of the dynamic space-time diagram topology, invoking a heterogeneous space-time data set according to the heterogeneous electricity price driving factor set, performing time sequence model training node by node, and constructing K core time sequence prediction models and Q auxiliary feature extraction models; And respectively extracting K core hidden state vectors and Q auxiliary hidden state vectors from the K core time sequence prediction models and the Q auxiliary feature extraction models to serve as initial node features, carrying out node state iterative updating of the dynamic space-time diagram topology by adopting a multi-layer space-time diagram convolution network, and outputting an electricity price prediction diagram neural network model.
4. The new energy station electricity price prediction optimization method based on multi-source fusion learning according to claim 2, wherein electricity price fluctuation association capturing is performed on new energy electricity price multielement influence factors to obtain the heterogeneous electricity price driving factor set, and the method comprises the following steps: Identifying and decomposing the new energy electricity price multielement influence factors based on factor modes to obtain a plurality of continuous factors and a plurality of discrete factors; Triggering fluctuation event dynamic slicing of the continuous factors according to the electricity price fluctuation node sequence to obtain a plurality of continuous associated time sequence fragments; Performing time lag mutual information calculation on the electricity price fluctuation node sequence and a plurality of continuous type associated time sequence fragments by adopting a sliding time window, and outputting a plurality of time lag associated strength values; Traversing the plurality of time-lag correlation strength values by adopting a preset core contribution threshold value to screen the K core driving factors; A plurality of event node sequences of the discrete factors are called, and a plurality of event impact strength values are output by calculating distribution KL divergences of the electricity price fluctuation node sequences and the event node sequences; Traversing the plurality of event impact strength values by adopting a preset auxiliary association threshold to screen the Q auxiliary driving factors.
5. The new energy station electricity price prediction optimization method based on multi-source fusion learning according to claim 4, wherein after setting K core prediction nodes and Q auxiliary feature nodes based on the K core driving factors and Q auxiliary driving factors respectively, performing node-by-node dynamic binding based on a dynamic edge relation rule to complete construction of the dynamic space-time diagram topology, the method comprises: enumerating the K core prediction nodes and the Q auxiliary feature nodes in a combined manner to obtain a plurality of node relation pairs; according to the node relation pairs, performing dynamic calling and splicing of the continuous associated time sequence fragments and the event node sequences to obtain a plurality of groups of node pair combined characteristic sequences; Performing time-lag mutual information calculation on the multi-group node pair combination characteristic sequences by adopting a sliding time window, and outputting a plurality of node pair association strength values; Projecting the K core prediction nodes and the Q auxiliary feature nodes to a power grid GIS topological graph to perform physical adjacency pruning, and constructing a basic physical topology; and traversing the association strength values of the plurality of node pairs by adopting the dynamic edge relationship rule, carrying out dynamic association edge binding of the basic physical topology, and outputting the dynamic space-time diagram topology.
6. The new energy station electricity price prediction optimization method based on multi-source fusion learning according to claim 3, wherein after distributing LSTM units to K core prediction nodes and Q auxiliary feature nodes of the dynamic space-time diagram topology, a heterogeneous space-time data set is called according to the heterogeneous electricity price driving factor set, a time sequence model training is performed node by node, and K core time sequence prediction models and Q auxiliary feature extraction models are constructed, the method comprises: distributing double-layer gating LSTM units to the K core prediction nodes; distributing lightweight LSTM units to the Q assist feature nodes; k core heterogeneous time sequence data are called from the heterogeneous electricity price driving factor set, rolling prediction online training is carried out on the K core prediction nodes, and the K core time sequence prediction models are output, wherein the K core time sequence prediction models are updated online in minute level; And Q auxiliary heterogeneous intermittent data are fetched from the heterogeneous electricity price driving factor set, event-triggered comparison feature learning is conducted on the Q auxiliary feature nodes, and the Q auxiliary feature extraction models are output, wherein the Q auxiliary feature extraction models are updated based on event-driven intermittent data.
7. The new energy station electricity price prediction optimization method based on multi-source fusion learning as claimed in claim 3, wherein K core hidden state vectors and Q auxiliary hidden state vectors are extracted from the K core time sequence prediction models and Q auxiliary feature extraction models respectively as initial node features, and a multi-layer space-time diagram convolutional network is adopted to perform node state iterative update of the dynamic space-time diagram topology, and an electricity price prediction diagram neural network model is output, and the method comprises: extracting hidden states and prediction error characteristics of the K core time sequence prediction models, and generating K core hidden state vectors; Extracting output layer activation values of the Q auxiliary feature extraction models to generate Q auxiliary implicit state vectors; Projecting the K core hidden state vectors and the Q auxiliary hidden state vectors to a unified feature space to construct a multidimensional node feature matrix; In the process of executing node state iterative updating of the multi-layer space-time diagram convolution network based on the multi-dimensional node feature matrix and the adjacency matrix of the dynamic space-time diagram topology, performing evolution feature capturing of the multi-dimensional node feature matrix through causal convolution to obtain a time enhancement feature matrix; Inputting the time enhancement feature matrix into a space diagram convolution layer of the multi-layer space-time diagram convolution network, aggregating topology neighborhood information by taking the adjacent matrix as weight, and generating an updated node feature matrix through gate residual connection; Inputting the updated node feature matrix into a full connection layer of the multi-layer space-time diagram convolution network, and separating quantile electricity price prediction and K updated core feature vectors; and iterating until the KL divergences of the K updated core feature vectors and the K core hidden state vectors meet a convergence threshold, fixing network parameters, and outputting the electricity price prediction graph neural network model.
8. The new energy station electricity price prediction optimization method based on multi-source fusion learning according to claim 1, wherein the k+q dynamic feature coding vectors are loaded to K core prediction nodes and Q auxiliary feature nodes of the electricity price prediction graph neural network model, the topology co-evolution reasoning of the electricity price prediction graph is performed, and an electricity price optimization prediction result is output, and the method comprises: loading the K+Q dynamic feature coding vectors to K core prediction nodes and Q auxiliary feature nodes of the electricity price prediction graph neural network model, and executing node level fraction prediction to output K+Q initial prediction vectors; And taking the K+Q initial prediction vectors as node characteristic evolution inputs, executing the topological collaborative evolution reasoning of the electricity price prediction graph on the electricity price prediction graph neural network model, and outputting the electricity price optimization prediction result.
9. The new energy station electricity price prediction optimization method based on multi-source fusion learning according to claim 5, wherein each node relation pair is any one of a core-core node pair, a core-auxiliary node pair or an auxiliary-auxiliary node pair.
10. The new energy station electricity price prediction optimization system based on multi-source fusion learning is characterized by being used for implementing the new energy station electricity price prediction optimization method based on multi-source fusion learning as claimed in any one of claims 1-9, and the system comprises: the topology construction module is used for constructing a dynamic space-time diagram topology based on the heterogeneous electricity price driving factor set; The iteration updating module is used for carrying out node state iteration updating on the dynamic space-time diagram topology through a hierarchical collaborative learning mechanism after distributing the LSTM units to the K core prediction nodes and the Q auxiliary feature nodes of the dynamic space-time diagram topology, and outputting an electricity price prediction diagram neural network model; The fusion coding module is used for carrying out heterogeneous mode feature fusion coding on the K+Q real-time driving factor data according to the K+Q attribute embedded characterizations to obtain K+Q dynamic feature coding vectors; The evolution reasoning module is used for loading the K+Q dynamic feature coding vectors to K core prediction nodes and Q auxiliary feature nodes of the electricity price prediction graph neural network model, executing topology collaborative evolution reasoning of the electricity price prediction graph and outputting electricity price optimization prediction results, wherein the electricity price optimization prediction results comprise electricity price prediction values and uncertainty quantization intervals; and the risk control module is used for sending the electricity price optimization prediction result to a new energy station transaction decision-making end, and generating and controlling the risk by referring to the electric power market bidding strategy.

Description

New energy station electricity price prediction optimization method and system based on multisource fusion learning Technical Field The invention relates to the technical field of electricity price prediction, in particular to a new energy station electricity price prediction optimization method and system based on multi-source fusion learning. Background With the gradual increase of the generation proportion of new energy, especially the wide application in the form of energy with larger fluctuation such as wind energy, solar energy and the like, the electricity price prediction of the electric power market becomes more and more complex. Conventional electricity price prediction models typically rely on a single data source, such as historical electricity price, single electricity demand data, etc., ignoring external factors such as weather, complexity of supply and demand structure. Fluctuations in electricity prices are the result of the co-action of multiple factors, the interaction between which is complex and difficult to capture by a single data source. Therefore, the prediction method based on the single data source is not good in the electricity price prediction of the new energy power generation station, and cannot provide an accurate prediction result. Disclosure of Invention The application provides a new energy station electricity price prediction optimization method and system based on multi-source fusion learning, and aims to solve the technical problem that in the prior art, a prediction model with a single data source is generally adopted, and the correlation among various driving factors cannot be effectively captured in a complex and changeable electric power market environment, so that the accuracy of a prediction result is insufficient. The application discloses a first aspect of a new energy station electricity price prediction optimization method based on multi-source fusion learning, which comprises the steps of constructing a dynamic space-time diagram topology based on a heterogeneous electricity price driving factor set, distributing an LSTM unit to K core prediction nodes and Q auxiliary feature nodes of the dynamic space-time diagram topology, carrying out node state iterative updating of the dynamic space-time diagram topology through a hierarchical collaborative learning mechanism, outputting an electricity price prediction diagram neural network model, carrying out heterogeneous mode feature fusion coding on K+Q real-time driving factor data according to K+Q attribute embedded characterizations to obtain K+Q dynamic feature coding vectors, loading the K+Q dynamic feature coding vectors to K core prediction nodes and Q auxiliary feature nodes of the electricity price prediction diagram neural network model, carrying out electricity price prediction diagram topology collaborative evolution reasoning, outputting electricity price optimization prediction results, wherein the electricity price optimization prediction results comprise electricity price prediction values and uncertainty quantization intervals, and sending the electricity price optimization prediction results to a new energy station trading end, and carrying out reference risk competition strategy generation and competition control. The application discloses a second aspect, provides a new energy station electricity price prediction optimization system based on multi-source fusion learning, which is used for the new energy station electricity price prediction optimization method based on multi-source fusion learning, and comprises a topology construction module, an iteration update module, an evolution inference module and a decision-making module, wherein the topology construction module is used for constructing a dynamic space-time diagram topology based on a heterogeneous electricity price driving factor set, the iteration update module is used for distributing an LSTM unit to K core prediction nodes and Q auxiliary feature nodes of the dynamic space-time diagram topology, performing node state iteration update of the dynamic space-time diagram topology through a hierarchical collaborative learning mechanism and outputting an electricity price prediction diagram neural network model, the fusion encoding module is used for performing heterogeneous modal feature fusion encoding on the K+Q real-time driving factor data according to K+Q attribute embedded characterizations to obtain K+Q dynamic feature encoding vectors, the evolution inference module is used for loading the K+Q dynamic feature encoding vectors to the K core prediction nodes and Q auxiliary feature nodes of the electricity price prediction diagram neural network model, performing topology collaborative evolution of the electricity price prediction diagram, and outputting an electricity price prediction diagram, wherein the price prediction optimization result comprises a price prediction result and a risk prediction risk control and a risk optimization