CN-122000873-A - Intelligent prediction method, system and storage medium for power load

CN122000873ACN 122000873 ACN122000873 ACN 122000873ACN-122000873-A

Abstract

The application relates to the technical field of electricity load prediction and discloses an intelligent electricity load prediction method, an intelligent electricity load prediction system and a storage medium. The method comprises the steps of carrying out regional difference quantification on load data of a source domain and target domain to obtain regional feature vectors, carrying out time sequence coding and attention calculation on the regional feature vectors to obtain time sequence features and weight distribution of the load of the source domain, obtaining domain invariant load features through domain countermeasure training, carrying out domain crossing migration by combining knowledge fusion to obtain the load features of the target domain, and generating a cross-regional load prediction result through a multi-step prediction algorithm after fine adjustment of historical data of the target domain is fused. The method solves the problem of prediction accuracy reduction caused by the characteristic distribution difference of the source domain and the target domain in cross-regional load prediction.

Inventors

WANG HONGGANG
WANG CHAO
ZHANG JIAN
JI HONGGEN
WANG SEN
WANG XIAOHUI
LI SHISHUAI
WANG GUOHUA
WANG XINZHAO

Assignees

河南泰隆电力设备股份有限公司

Dates

Publication Date: 20260508
Application Date: 20260115

Claims (10)

1. An intelligent prediction method for electricity load, which is characterized by comprising the following steps: carrying out regional difference quantization processing on the source domain and target domain load data through a regional load pattern recognition algorithm to obtain a regional characteristic vector containing a load rate, a peak Gu Chalv and a load fluctuation coefficient; performing time sequence coding and attention weight calculation processing on the regional feature vector to obtain source domain load time sequence features and load attention weight distribution; performing domain adaptation processing on the source domain load time sequence characteristics through domain countermeasure training to obtain domain invariant load characteristics; performing cross-domain migration processing on the domain invariant load characteristics and the load attention weight distribution according to a knowledge fusion algorithm to obtain target domain load characteristics; and carrying out fusion fine tuning processing on the target domain historical load data and the target domain load characteristics to obtain optimized target domain load characteristics, and carrying out time sequence prediction processing on the optimized target domain load characteristics according to a multi-step prediction algorithm to generate a cross-region load prediction result.
2. The intelligent prediction method of electric load according to claim 1, wherein the performing the regional differential quantization processing on the source domain and the target domain load data by using the regional load pattern recognition algorithm to obtain a regional feature vector including a load factor, a peak Gu Chalv and a load fluctuation coefficient comprises: performing outlier detection and missing value interpolation processing on historical power load data of a source domain and a target domain to obtain standardized load time sequence data; Calculating an average load value and a maximum load value based on the standardized load time sequence data, and carrying out ratio calculation processing on the average load value and the maximum load value to obtain a load rate value; Extracting peak load and valley load based on the normalized load time sequence data, dividing the difference value of the peak load and the valley load by the average load value, and carrying out peak-valley difference rate calculation processing to obtain a peak Gu Chalv value; Calculating a load standard deviation based on the normalized load time sequence data, carrying out ratio calculation processing on the load standard deviation and an average load value to obtain a load fluctuation coefficient value, and carrying out vector combination processing on the load rate, the peak Gu Chalv and the load fluctuation coefficient to obtain a regional characteristic vector.
3. The method for intelligently predicting the power consumption load according to claim 1, wherein the step of performing time sequence coding and attention weight calculation processing on the regional feature vector to obtain source domain load time sequence features and load attention weight distribution comprises the following steps: Carrying out sequence segmentation processing on the regional feature vectors according to a time sequence through a sliding window mechanism to obtain a load feature sequence segment with a fixed length; Performing multi-scale time sequence decomposition processing on the load characteristic sequence segment, respectively extracting a short-term fluctuation component, a daily period component and a long-term trend component, and performing characteristic fusion processing on the three components to obtain multi-scale time sequence characterization; based on the multi-scale time sequence representation, generating a position coding vector according to the relative position of each time step by a position encoder, and performing superposition processing on the position coding vector and the multi-scale time sequence representation to obtain a source domain load time sequence feature; Calculating the relevance scores among different time steps by the source domain load time sequence characteristics through a self-attention mechanism to obtain a time sequence relevance matrix; And carrying out normalization processing on the time sequence correlation matrix to ensure that the sum of the attention weights of each time step is a unit value, and obtaining the load attention weight distribution.
4. The intelligent prediction method of power consumption load according to claim 1, wherein the performing domain adaptation processing on the source domain load time sequence characteristic through domain countermeasure training to obtain a domain invariant load characteristic comprises: Inputting the source domain load time sequence characteristics into a load characteristic extraction network to perform deep characteristic extraction processing, and inputting the regional load data of the target domain into the same load characteristic extraction network to perform characteristic extraction processing to obtain source domain load deep characteristics and target domain load deep characteristics; Based on the load deep features of the source domain and the load deep features of the target domain, carrying out load source region classification processing through a region discriminator, calculating a region classification loss value, and carrying out counter propagation processing on gradients of region classification loss through a gradient inversion layer to obtain an inversion gradient signal for confusing region differences; Updating parameters of the load feature extraction network according to the inversion gradient signals, so that the load features generated by the load feature extraction network cannot be identified by the region identifier to obtain region-independent load confusion features; And carrying out inter-region load mode distance calculation processing on the region-independent load confusion characteristic through inter-region load distribution difference measurement, and stopping training when the inter-region load mode distance is smaller than a preset convergence threshold value to obtain a region-invariant load characteristic.
5. The intelligent prediction method of electric load according to claim 1, wherein the performing cross-domain migration processing on the domain invariant load feature and the load attention weight distribution according to the knowledge fusion algorithm to obtain a target domain load feature comprises: Carrying out characteristic compression processing on the domain invariant load characteristics through a knowledge distillation network to obtain compressed common load knowledge; Weighting selection processing is carried out on the common load knowledge based on the load attention weight distribution, and knowledge segments with higher relevance to the load mode of the target domain are screened out according to the attention weight value to obtain the load knowledge related to the target domain; carrying out knowledge fusion processing on the target domain related load knowledge and the regional feature vector of the target domain, and dynamically adjusting the fusion proportion of the source domain knowledge and the target domain feature through a gating mechanism to obtain a preliminary target domain load feature; And carrying out feature reconstruction processing on the preliminary target domain load features through a region specificity recovery network, and adjusting the features according to the industrial structure and electricity utilization habit of the target domain to obtain the target domain load features.
6. The intelligent prediction method of electricity load according to claim 1, wherein the fusing and fine-tuning the historical load data of the target domain and the load characteristic of the target domain to obtain the load characteristic of the target domain after optimization, performing time sequence prediction processing on the load characteristic of the target domain after optimization according to a multi-step prediction algorithm, and generating a cross-region load prediction result, includes: performing time sequence matching processing on the historical load data of the target domain and the load characteristics of the target domain through a time alignment mechanism to obtain time aligned load data characteristic pairs; based on the time-aligned load data feature pairs, carrying out fusion processing on local mode information of the historical load data of the target domain and the load features of the target domain through a residual error connection network, and dynamically adjusting contribution weights of the historical data and the migration features to obtain fused load features of the target domain; Carrying out region specificity adjustment processing on the fused target domain load characteristics through a parameter fine adjustment network, and carrying out fine granularity correction on the characteristics according to the actual load change rule of the target domain to obtain optimized target domain load characteristics; And inputting the optimized target domain load characteristics into a multi-step time sequence prediction network to carry out recursive prediction processing, wherein each step of prediction result is used as the input characteristic of the next step, and error propagation of multi-step prediction is restrained through an error accumulation control mechanism, so that a trans-regional load prediction result is obtained.
7. The intelligent prediction method of electric load according to claim 6, wherein the performing region-specific adjustment processing on the fused target domain load characteristic through a parameter fine adjustment network, performing fine-grained correction on the characteristic according to an actual load change rule of the target domain, and obtaining an optimized target domain load characteristic includes: based on the fused load characteristics of the target domain, extracting load change modes of the target domain in the early peak, the late peak, the midnight low valley and the late night low valley period of the working day, and calculating load average values and load change rates of all periods to obtain load mode parameters of the target domain period; According to the target domain time period load mode parameters, converting the load characteristics of each time period into weight adjustment coefficients of a fine adjustment network through a parameter mapping function, wherein the value range of the daytime production time period adjustment coefficients of an industrial area is 1.2-1.5, and the value range of the early and late electricity consumption peak adjustment coefficients of a residential area is 1.3-1.6, so that regional weight adjustment coefficients are obtained; inputting the fused target domain load characteristics and the regional weight adjustment coefficients into a parameter fine adjustment network to perform characteristic correction processing, and performing differential adjustment on characteristic values of different time periods according to the actual industrial structure proportion of the target domain and the electricity utilization habit of residents to obtain time period differential load characteristics; and carrying out temperature sensitivity correction processing on the time period differentiated load characteristics through a seasonal correction function, and adjusting the characteristics of the corresponding seasons according to the historical data of the summer refrigeration load and the winter heating load of the target domain to obtain the optimized load characteristics of the target domain.
8. An intelligent electricity load prediction system for implementing the intelligent electricity load prediction method according to any one of claims 1 to 7, comprising: the quantization module is used for carrying out regional difference quantization processing on the source domain and target domain load data through a regional load pattern recognition algorithm to obtain a regional characteristic vector containing a load rate, a peak Gu Chalv and a load fluctuation coefficient; the computing module is used for carrying out time sequence coding and attention weight computing processing on the regional characteristic vector to obtain source domain load time sequence characteristics and load attention weight distribution; the processing module is used for carrying out domain adaptation processing on the source domain load time sequence characteristics through domain countermeasure training to obtain domain invariant load characteristics; the migration module is used for carrying out cross-domain migration processing on the domain invariant load characteristics and the load attention weight distribution according to a knowledge fusion algorithm to obtain target domain load characteristics; The prediction module is used for carrying out fusion fine adjustment processing on the target domain historical load data and the target domain load characteristics to obtain optimized target domain load characteristics, carrying out time sequence prediction processing on the optimized target domain load characteristics according to a multi-step prediction algorithm, and generating a cross-region load prediction result.
9. An intelligent electricity load prediction device comprising a memory and a processor, the memory storing a computer program executable on the processor, the processor implementing the intelligent electricity load prediction method of any one of claims 1 to 7 when the computer program is executed by the processor.
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, causes the processor to perform the electrical load intelligent prediction method according to any one of claims 1 to 7.

Description

Intelligent prediction method, system and storage medium for power load Technical Field The application relates to the technical field of electricity load prediction, in particular to an intelligent electricity load prediction method, an intelligent electricity load prediction system and a storage medium. Background The existing electricity load prediction technology is mainly used for modeling based on historical data of a single area, and predicting the electricity load of a local area through methods such as time sequence analysis and machine learning. These methods generally require a large amount of historical load data for model training, and can achieve a good prediction effect for a region with sufficient data. However, the prior art has obvious defects when facing a new area or a target area with scarce historical data, on one hand, the new area lacks enough historical data to support model training, so that the prediction precision is low, and on the other hand, when the models trained in other areas are directly applied to the target area, the prediction result tends to deviate greatly due to differences of industrial structures, electricity utilization habits, climate conditions and the like of different areas. In addition, the existing method lacks an effective trans-regional knowledge migration mechanism, and cannot fully utilize the prediction experience of the data-rich region. Even if a simple migration learning method is adopted, the progressive technical problems still exist that firstly, the load characteristic distribution difference between a source domain and a target domain is difficult to effectively eliminate, so that migrated characteristics comprise region specific deviation, secondly, a targeted knowledge screening mechanism is lacked, which source domain knowledge is really useful for target domain prediction cannot be identified, and finally, a migrated model is lacked fusion adjustment with the actual condition of the target domain, and the local characteristics of the target domain are ignored. Disclosure of Invention The application provides an intelligent prediction method, system and storage medium for power loads, which are used for solving the problem of prediction accuracy reduction caused by characteristic distribution difference of source domains and target domains in cross-regional load prediction, further solving the problem of poor migration effect caused by lack of effective knowledge screening and fusion mechanisms even after the distribution difference is eliminated, and finally solving the problem that an accurate prediction model cannot be built in a newly built or data scarce area due to lack of enough historical data. The application provides an intelligent prediction method of electric loads, which comprises the steps of carrying out regional difference quantification processing on source domain and target domain load data through a regional load pattern recognition algorithm to obtain regional characteristic vectors containing load rates, peaks Gu Chalv and load fluctuation coefficients, carrying out time sequence coding and attention weight calculation processing on the regional characteristic vectors to obtain source domain load time sequence characteristics and load attention weight distribution, carrying out domain adaptation processing on the source domain load time sequence characteristics through domain countermeasure training to obtain domain invariant load characteristics, carrying out cross-domain migration processing on the domain invariant load characteristics and the load attention weight distribution according to a knowledge fusion algorithm to obtain target domain load characteristics, carrying out fusion fine adjustment processing on target domain historical load data and the target domain load characteristics to obtain optimized target domain load characteristics, carrying out time sequence prediction processing on the optimized target domain load characteristics according to a multi-step prediction algorithm, and generating a cross-region load prediction result. Optionally, the performing a region difference quantization process on the load data of the source domain and the target domain by using a region load pattern recognition algorithm to obtain a region feature vector including a load rate, a peak Gu Chalv and a load fluctuation coefficient, including: performing outlier detection and missing value interpolation processing on historical power load data of a source domain and a target domain to obtain standardized load time sequence data; Calculating an average load value and a maximum load value based on the standardized load time sequence data, and carrying out ratio calculation processing on the average load value and the maximum load value to obtain a load rate value; Extracting peak load and valley load based on the normalized load time sequence data, dividing the difference value of the peak load and the valley load by the aver