CN-122021995-A - Load prediction method based on dual feature processing and error correction

CN122021995ACN 122021995 ACN122021995 ACN 122021995ACN-122021995-A

Abstract

The invention relates to the technical field of machine learning, and discloses a load prediction method based on dual feature processing and error correction. The method comprises the steps of obtaining multi-source time sequence data, decomposing historical load data into a plurality of modal components by adopting a variation modal decomposition algorithm, classifying each modal component into different frequency grades according to sample entropy, carrying out phase space reconstruction according to modal components of each frequency grade and corresponding external influence factor data to generate multivariable phase space data sets of each frequency grade, respectively inputting the multivariable phase space data sets of each frequency grade into a corresponding load prediction sub-model to generate a prediction output result and superposing the prediction output result, constructing a residual sequence based on the historical load data and an initial load prediction result, inputting the residual sequence into a residual prediction model to obtain a load residual prediction value, and compensating the initial load prediction result through the load residual prediction value. According to the scheme, the load prediction accuracy can be improved.

Inventors

XUE DONG
XU JINGJING
SU JIANFENG
JIANG TING
XU CONG
YAN LICHEN
WANG LIQIANG
Liu Zhaoxinyu

Assignees

中国华电科工集团有限公司
华电综合智慧能源科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251219

Claims (10)

1. A load prediction method based on dual feature processing and error correction, the method comprising: The method comprises the steps of obtaining multi-source time sequence data of a target area, wherein the multi-source time sequence data comprises historical load data and corresponding external influence factor data; Decomposing the historical load data into a plurality of modal components by adopting a variational modal decomposition algorithm; sample entropy of each modal component is calculated respectively, and each modal component is classified into different frequency grades according to the sample entropy; carrying out phase space reconstruction according to the modal components of each frequency level and the corresponding external influence factor data to generate a multivariable phase space data set of each frequency level; Respectively inputting the multivariable phase space data set of each frequency level into a corresponding load prediction sub-model, generating a prediction output result of each frequency level, and superposing the prediction output results to obtain an initial load prediction result; Constructing a residual sequence based on the historical load data and the initial load prediction result; inputting the residual sequence into a residual prediction model to obtain a load residual prediction value; and compensating the initial load prediction result through the load residual error prediction value to obtain a load prediction result.
2. The method of claim 1, wherein decomposing the historical load data into a plurality of modal components using a variational modal decomposition algorithm comprises: Decomposing the historical load data into modal components of target modal numbers based on an optimal penalty factor and an optimal Lagrange multiplier by adopting a variable modal decomposition algorithm; the method further comprises the steps of: setting a mode number candidate range and a penalty factor value interval in a training stage; traversing candidate parameter combinations consisting of any modal number in a modal number candidate range and any penalty factor in a penalty factor value interval, and executing the following variation modal decomposition process on sample historical load data in a training set: Establishing a constrained variation optimization model by taking the sum of the modal components as constraint conditions and the minimum sum of the estimated bandwidths of the modal components as a target; Converting the constrained variational optimization model into an augmented Lagrangian function by introducing penalty factors and Lagrangian multipliers; initializing the center frequency of the target number and the corresponding modal components by taking the modal number in the corresponding candidate parameter combination as the target number; Iteratively updating the augmented Lagrangian function by a multiplication operator alternating direction method, and synchronously optimizing each modal component and the corresponding center frequency; When the change between the modal components of two adjacent iterations is smaller than a preset threshold, judging convergence, and outputting the modal components of the target number; repeating the variation modal decomposition process to obtain a modal component set of each candidate parameter combination; And evaluating the modal component set of each candidate parameter combination, and selecting the candidate parameter combination with the optimal evaluation result to determine the optimal penalty factor and the optimal Lagrange multiplier.
3. The method according to claim 1, wherein calculating sample entropy of each modal component and classifying each modal component into different frequency classes according to sample entropy size comprises: respectively calculating the sample entropy of each modal component; Comparing the sample entropy of each modal component with a preset sample entropy threshold value, and dividing each modal component into different frequency levels according to the comparison result.
4. The method of claim 1, wherein the phase space reconstruction from the modal components of each frequency level and the corresponding external influence factor data to generate a multi-variable phase space dataset of each frequency level comprises: Based on the optimal delay time and the optimal embedding dimension, performing phase space reconstruction according to the modal component of each frequency level and corresponding external influence factor data, and generating a multivariable phase space data set of each frequency level; the method further comprises the steps of: In the training stage, constructing a candidate delay time set; according to the first leachable weight vector, weighting and fusing each candidate delay time in the candidate delay time set to generate a delay component; constructing multidimensional embedded vectors of the phase space based on the delay components, and carrying out weighting processing on each embedded dimension of the phase space according to the second learnable weight vector; presetting a maximum embedding dimension, and sequentially attenuating each embedding dimension according to the contribution degree to a prediction task in the training process until a preset sparse condition is met to obtain an effective embedding dimension; Constructing a total loss function based on the phase space reconstruction coincidence loss and the load prediction sub-model loss; synchronously updating parameters of the first and second learnable weight vectors and the load prediction sub-model by using a back propagation algorithm with the aim of minimizing the total loss function; When the training of the load prediction sub-model converges, determining the optimal delay time according to the final first learnable weight vector, and determining the optimal embedding dimension according to the final second learnable weight vector.
5. The method according to any one of claims 1 to 4, wherein the inputting the multivariate phase space dataset of each frequency class into the corresponding load predictor model, respectively, generates a predicted output result of each frequency class comprises: extracting long-term time dependent characteristics of a multivariable phase space dataset corresponding to the frequency level through a long-term memory network layer of the load prediction sub-model to obtain first hidden layer output at each moment; distributing weights to hidden layer outputs at all moments through an attention mechanism layer of the load prediction sub-model, and carrying out weighted summation to obtain a first context vector; And carrying out nonlinear transformation on the first context vector through an output layer of the load prediction sub-model to generate a prediction output result corresponding to the frequency level.
6. The method of claim 5, wherein inputting the residual sequence into a residual prediction model to obtain a load residual prediction value comprises: extracting long-term time dependent features of a residual sequence through a long-term and short-term memory network layer of a residual prediction model, and generating a second hidden layer output at each moment; Distributing weights to the second hidden layer output at each moment through the attention mechanism layer of the residual prediction model, and carrying out weighted summation to obtain a second context vector; and carrying out nonlinear transformation on the second context vector through an output layer of the residual prediction model to generate a load residual prediction value.
7. A load predicting device based on dual feature processing and error correction, the device comprising: The acquisition module is used for acquiring multi-source time sequence data of the target area, wherein the multi-source time sequence data comprises historical load data and corresponding external influence factor data; the modal decomposition module is used for decomposing the historical load data into a plurality of modal components by adopting a variation modal decomposition algorithm; the frequency grade module is used for respectively calculating sample entropy of each modal component and classifying each modal component into different frequency grades according to the sample entropy; The phase space reconstruction module is used for carrying out phase space reconstruction according to the modal component of each frequency level and the corresponding external influence factor data to generate a multivariable phase space data set of each frequency level; the load prediction module is used for respectively inputting the multivariable phase space data set of each frequency level into the corresponding load prediction sub-model, generating a prediction output result of each frequency level and superposing the prediction output results to obtain an initial load prediction result; the residual sequence construction module is used for constructing a residual sequence based on the historical load data and the initial load prediction result; the load residual prediction module is used for inputting the residual sequence into a residual prediction model to obtain a load residual prediction value; And the compensation module is used for compensating the initial load prediction result through the load residual error prediction value to obtain a load prediction result.
8. An electronic device, comprising: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory having stored therein computer instructions, the processor executing the computer instructions to perform the dual feature processing and error correction based load prediction method of any one of claims 1 to 6.
9. A computer-readable storage medium having stored thereon computer instructions for causing a computer to execute the dual feature processing and error correction-based load prediction method according to any one of claims 1 to 6.
10. A computer program product comprising computer instructions for causing a computer to perform the dual feature processing and error correction based load prediction method of any one of claims 1 to 6.

Description

Load prediction method based on dual feature processing and error correction Technical Field The invention relates to the technical field of machine learning, in particular to a load prediction method based on dual feature processing and error correction. Background In order to improve the operational energy efficiency of the energy system, an accurate and reliable prediction of the thermal load is required. Because the integrated energy system involves the mutual coupling between different energy flows, the heating load often presents complex variation characteristics, and brings great challenges to the optimized operation of the system. In the related art, a time sequence method, a regression analysis method, a support vector machine, an artificial neural network and the like are adopted for load prediction, however, the methods are difficult to comprehensively and effectively extract and utilize complex characteristic information in a heating load, or problems such as characteristic redundancy, overfitting and the like easily occur, and an error correction mechanism is lacked, so that the prediction accuracy is poor. Disclosure of Invention The invention provides a load prediction method based on dual feature processing and error correction, which aims to solve the problem of poor accuracy in load prediction in the related art. In a first aspect, the present invention provides a load prediction method based on dual feature processing and error correction, the method comprising: The method comprises the steps of obtaining multi-source time sequence data of a target area, wherein the multi-source time sequence data comprises historical load data and corresponding external influence factor data; Decomposing the historical load data into a plurality of modal components by adopting a variational modal decomposition algorithm; sample entropy of each modal component is calculated respectively, and each modal component is classified into different frequency grades according to the sample entropy; carrying out phase space reconstruction according to the modal components of each frequency level and the corresponding external influence factor data to generate a multivariable phase space data set of each frequency level; Respectively inputting the multivariable phase space data set of each frequency level into a corresponding load prediction sub-model, generating a prediction output result of each frequency level, and superposing the prediction output results to obtain an initial load prediction result; Constructing a residual sequence based on the historical load data and the initial load prediction result; inputting the residual sequence into a residual prediction model to obtain a load residual prediction value; and compensating the initial load prediction result through the load residual error prediction value to obtain a load prediction result. In an alternative embodiment, the decomposing the historical load data into a plurality of modal components using a variational modal decomposition algorithm includes: Decomposing the historical load data into modal components of target modal numbers based on an optimal penalty factor and an optimal Lagrange multiplier by adopting a variable modal decomposition algorithm; the method further comprises the steps of: setting a mode number candidate range and a penalty factor value interval in a training stage; traversing candidate parameter combinations consisting of any modal number in a modal number candidate range and any penalty factor in a penalty factor value interval, and executing the following variation modal decomposition process on sample historical load data in a training set: Establishing a constrained variation optimization model by taking the sum of the modal components as constraint conditions and the minimum sum of the estimated bandwidths of the modal components as a target; Converting the constrained variational optimization model into an augmented Lagrangian function by introducing penalty factors and Lagrangian multipliers; initializing the center frequency of the target number and the corresponding modal components by taking the modal number in the corresponding candidate parameter combination as the target number; Iteratively updating the augmented Lagrangian function by a multiplication operator alternating direction method, and synchronously optimizing each modal component and the corresponding center frequency; When the change between the modal components of two adjacent iterations is smaller than a preset threshold, judging convergence, and outputting the modal components of the target number; repeating the variation modal decomposition process to obtain a modal component set of each candidate parameter combination; And evaluating the modal component set of each candidate parameter combination, and selecting the candidate parameter combination with the optimal evaluation result to determine the optimal penalty factor and the optimal Lagrange multi