CN-122022003-A - Method and device for predicting carbon emission of generator set on power generation side based on time sequence large model

CN122022003ACN 122022003 ACN122022003 ACN 122022003ACN-122022003-A

Abstract

The invention relates to the technical field of carbon emission prediction, in particular to a method and a device for predicting carbon emission of a generator set on the side of a time sequence large model, which comprise the steps of determining a time sequence input sequence of a pre-trained time sequence large model based on multi-source basic data of a power station; and taking the time sequence input sequence as the input of a pre-trained time sequence large model to obtain a power station carbon emission prediction result output by the pre-trained time sequence large model, wherein the multi-source basic data comprise power station load data, carbon emission grid data and ERA5 meteorological data. According to the technical scheme provided by the invention, the spatial characteristics of the spatial data can be effectively extracted and fused, and the high-precision prediction is realized based on the limited data.

Inventors

SONG JINWEI
LI XIAOYAN
ZHANG SHIZE
CHEN XIANG
YAN YUE
Lou Dejun
LIU SHANSHAN
HE QI
BAI HAOTIAN
WANG CHUNMEI

Assignees

国家电网有限公司信息通信中心

Dates

Publication Date: 20260512
Application Date: 20251225

Claims (10)

1. The method for predicting the carbon emission of the generator set based on the time sequence large model is characterized by comprising the following steps of: determining a time sequence input sequence of a pre-trained time sequence large model based on multi-source basic data of a power plant station; Taking the time sequence input sequence as the input of a pre-trained time sequence large model to obtain a power plant station carbon emission prediction result output by the pre-trained time sequence large model; the multi-source basic data comprise power station load data, carbon emission grid data and ERA5 meteorological data.
2. The method of claim 1, wherein the determining a time series input sequence of a pre-trained time series large model based on multi-source base data of a power plant station comprises: performing space-time scale alignment on multi-source basic data of the power plant station; taking longitude and latitude of a power station as a center, extracting preset far-field scale grid data and preset near-field scale grid data; respectively carrying out feature extraction on the preset far-field scale grid data and the preset near-field scale grid data through a far-field encoder structure and a near-field encoder structure to obtain far-field grid data features and near-field grid data features; splicing the far-field grid data features and the near-field grid data features to obtain grid data features; Carrying out feature fusion on grid data features corresponding to carbon emission grid data and grid data features corresponding to ERA5 meteorological data by using an attention mechanism and a fully connected network to obtain fusion features; and splicing the time sequence corresponding to the fusion characteristic with the time sequence corresponding to the power station load data to obtain a time sequence input sequence.
3. The method of claim 2, wherein the spatio-temporal scale alignment of the multi-source base data of the power plant station comprises: Upsampling or downsampling the ERA5 weather data to make its spatial resolution consistent with carbon emission grid data; and (3) daily aggregation of the load data of the power plant stations, so that the time scale of the load data is consistent with the carbon emission grid data and the ERA5 meteorological data.
4. The method of claim 3, wherein after upsampling the ERA5 weather data, interpolating the data for the point to be interpolated according to the following formula: In the above-mentioned method, the step of, For points to be interpolated Is used for the interpolation of the data of (a), Four vertex coordinates of the ERA5 weather data grid, Positions in ERA5 weather data grids, respectively ERA5 weather data of (a).
5. The method of claim 2, wherein the predetermined far-field scale grid data corresponds to a scale of 17 x 17 and the predetermined near-field scale grid data corresponds to a scale of 7 x 7.
6. The method of claim 2, wherein the far field encoder structure comprises a first convolution layer, a first ReLU activation layer, a first max-pooling layer, and a first ConvLSTM layer, wherein the near field encoder structure comprises a second convolution layer, a second ReLU activation layer, a second max-pooling layer, a third convolution layer, a second ConvLSTM layer, and a global average pooling layer; The first convolution layer comprises 32 3×3 convolution kernels, the step length of the first convolution layer is 1, the filling is 1, the number of hidden units of the first ConvLSTM layer is 64, the time step length is 7, the second convolution layer comprises 32 7×7 convolution kernels, the step length of the second convolution layer is 1, the filling is 3, the third convolution layer comprises 64 5×5 convolution kernels, the step length of the third convolution layer is 1, the filling is 2, and the number of hidden units of the second ConvLSTM layer is 128, and the time step length is 7.
7. The method of claim 2, wherein the fusion characteristics are as follows: In the above-mentioned method, the step of, In order to fuse the features of the features, For the function to be activated by the ReLU, The first parameter is fully connected for the first layer, As a first attention weight to be paid to, For the grid data characteristics corresponding to the carbon emission grid data, For the second attention weighting to be given, For the grid data characteristics corresponding to ERA5 weather data, The second parameter is fully connected for the first layer, The first parameter is fully connected for the second layer, The second parameter is fully connected for the second layer.
8. The method of claim 7, wherein the first attention weight and the second attention weight are as follows: In the above-mentioned method, the step of, For the first attention layer parameter corresponding to the carbon emission grid data, A second attention layer parameter corresponding to the carbon emission grid data, For the first attention layer parameter corresponding to ERA5 weather data, And a second attention layer parameter corresponding to ERA5 meteorological data.
9. The method of claim 1, wherein the training process of the pre-trained time series large model comprises: Determining a time sequence input sequence of the time sequence large model based on historical multi-source basic data of the power plant station; constructing training data by utilizing the time sequence input sequence and the historical carbon emission result of the power plant station; Training the time sequence large model by utilizing the training data to obtain the pre-trained time sequence large model; And when the pre-trained time sequence large model is subjected to fine adjustment aiming at a new power station, training the pre-trained time sequence large model by using training data corresponding to the new power station, and freezing model parameters of the pre-trained time sequence large model in the training process to only adjust the parameters of the bottleneck structure adapter.
10. A generator-side unit carbon emission prediction device based on a time series large model, the device comprising: a determining module for determining a time sequence input sequence of a pre-trained time sequence large model based on multi-source basic data of the power plant station; the prediction module is used for taking the time sequence input sequence as the input of a pre-trained time sequence large model to obtain a power plant station carbon emission prediction result output by the pre-trained time sequence large model; the multi-source basic data comprise power station load data, carbon emission grid data and ERA5 meteorological data.

Description

Method and device for predicting carbon emission of generator set on power generation side based on time sequence large model Technical Field The invention relates to the technical field of carbon emission prediction, in particular to a method and a device for predicting carbon emission of a generator set on the side of a time sequence large model. Background Along with huge energy resource consumption in the electric power field, the energy system is clean, low-carbon, safe and efficient. The main source of carbon emission in the electric power field is concentrated on a generator side unit, and in the background, the refined prediction of the carbon emission of the generator side unit has a positive effect on the formulation of a carbon emission strategy. Through investigation of the current state of carbon emission prediction, the traditional power generation side carbon emission prediction mainly has two technical paths, namely, a method based on emission source quantification-statistics accounting, a method for identifying emission source parameters such as fuel consumption, combustion efficiency and the like of a generator set, a method for calculating historical carbon emission by combining an IPCC (IPCC) accounting formula, a method for extrapolating and predicting by using a traditional statistical learning method such as an autoregressive integrated moving average model (ARIMA) and a Support Vector Machine (SVM), and a method based on data driving-simple multivariable coupling, and a method for attempting to combine meteorological data as a covariate and predicting by using a Convolutional Neural Network (CNN) or a long-short-term memory network (LSTM). However, such methods are typically limited to univariate time-series fitting or linear variable coupling, difficult to capture complex nonlinear dependencies between carbon emissions-meteorological-geographic-loads, and poor generalization capability-feature engineering and model structures need to be redesigned for new power generation sites. With the success of large pre-training models in the fields of Natural Language Processing (NLP) and computer vision, the field of time sequence prediction gradually forms a new paradigm of pre-training and fine tuning, namely, a time sequence base model (TSFMs) is pre-trained through massive heterogeneous data, general space-time knowledge is learned, and downstream tasks (such as load prediction and weather prediction) can be quickly adapted. For example, chronos (ChronosX), timeGPT and other time sequence large models can process multidimensional data and complex time dependence, but in a power generation side carbon emission prediction scene, the technical bottleneck exists in that the power generation side data has the characteristics of strong spatial correlation and high data heterogeneity (grid data+point data), and the existing TSFMs only inputs the multisource data as flat covariates and cannot utilize the inherent spatial correlation information, so that the prediction accuracy is limited. Although improved attempts at adapter injection have emerged in the field of time series large models (TSFMs), there are two major core limitations to such designs that do not fit into the generator-side unit carbon emission prediction scenario: 1. The adapter has the function limitation that only single type time sequence data is processed, and multi-source space features are not fused 2. The adapter structure and scene adaptation are limited by the fact that the design of a lightweight bottleneck is omitted, and the insertion layer selects the adaptation logic without space characteristics. In addition, due to the problems of scarcity, fragmentation and the like of carbon emission data of the generator set, how to realize high-precision prediction based on limited data becomes a key bottleneck in the application of the current technology. Disclosure of Invention In order to overcome the defects, the invention provides a method and a device for predicting carbon emission of a generator set on the side of a time sequence large model. In a first aspect, a method for predicting carbon emission of a generator set based on a time-series large model is provided, where the method for predicting carbon emission of a generator set based on a time-series large model includes: determining a time sequence input sequence of a pre-trained time sequence large model based on multi-source basic data of a power plant station; Taking the time sequence input sequence as the input of a pre-trained time sequence large model to obtain a power plant station carbon emission prediction result output by the pre-trained time sequence large model; the multi-source basic data comprise power station load data, carbon emission grid data and ERA5 meteorological data. Preferably, the determining the time sequence input sequence of the pre-trained time sequence big model based on the multi-source basic data of the power station comprises: perfor