Search

CN-122026324-A - Multi-dimensional attribute power consumer time-sharing electric quantity clustering analysis prediction method and system

CN122026324ACN 122026324 ACN122026324 ACN 122026324ACN-122026324-A

Abstract

The invention discloses a multi-dimensional attribute power user time-sharing electric quantity cluster analysis prediction method and a system, wherein the method comprises the steps of extracting multi-dimensional prediction input characteristics including time sequence characteristics, market characteristics and business characteristics from acquired data; invoking a typical electricity consumption mode clustering template, calculating the similarity between a user to be predicted and each clustering template based on multidimensional prediction input characteristics, and determining target clusters and corresponding typical electric quantity curves; the method comprises the steps of obtaining an initial 24-hour time-sharing electric quantity weight of a user to be predicted through a multi-branch collaborative prediction model, dynamically adjusting the initial weight by combining an electric power spot market price signal and a price sensitivity level of the user to be predicted, synchronously executing business rule verification and correction, and outputting an optimized time-sharing electric quantity weight and a layered interpretation report. According to the invention, the prediction precision and stability are effectively improved by fusing the model prediction and the typical electricity consumption mode through the clustering enhancement strategy, and the technical support is provided for optimizing the spot market reporting strategy of the electricity selling company.

Inventors

  • ZHANG CHUANMING
  • MO LING
  • Ling Chunjian

Assignees

  • 九州能源有限公司

Dates

Publication Date
20260512
Application Date
20260127

Claims (10)

  1. 1. A multi-dimensional attribute power user time-sharing electric quantity cluster analysis prediction method is characterized by comprising the following steps: collecting basic data, electric power spot market data and business data of a user to be predicted, and extracting multidimensional prediction input characteristics including time sequence characteristics, market characteristics and business characteristics; invoking a pre-trained typical power consumption mode clustering template, calculating the similarity between a user to be predicted and each clustering template based on multidimensional prediction input characteristics, and determining target clusters and corresponding typical electric quantity curves; inputting multidimensional prediction input characteristics and target clustering information into a multi-branch collaborative prediction model to predict, and obtaining an initial 24-hour time-sharing electric quantity weight of a user to be predicted; Combining the power spot market price signal and the price sensitivity level of the user to be predicted, dynamically adjusting the power weight of the initial 24-hour time-sharing power, and synchronously executing the check and correction of the business rule; And outputting the optimized time-sharing electric quantity weight according to the daily, daily and real-time multiple time scales, and synchronously generating a layered interpretation report containing characteristic contribution and service compliance.
  2. 2. The multi-dimensional attribute power consumer time-sharing electric quantity cluster analysis prediction method according to claim 1, wherein the typical electric consumption mode cluster template comprises: The multi-source training data of 3 years of collection history are proportionally divided into a clustering training set and a verification set; triple clustering training is carried out on the training set, wherein the triple clustering comprises a first heavy space-time clustering, a second heavy density clustering and a third heavy business clustering; taking the median sequence of each cluster as a typical electric quantity curve template of the cluster; The method comprises the steps of monitoring application conditions of all templates in a template database in real time, and triggering instant update when any condition is met, wherein the change of a user sample when the actual user sample corresponding to any template is more than 15% compared with the template generation, or the actual electricity consumption data of any template continuously for 15 days exceeds 12% compared with MAPE of the template in the prediction application.
  3. 3. The multi-dimensional attribute power consumer time-sharing power cluster analysis prediction method according to claim 2, wherein the typical power consumption mode cluster template further comprises: The first weight space-time clustering adopts a space-time density clustering algorithm, a characteristic sequence of 7 continuous days of a single user is taken as 1 space-time sample, key parameters of ST-DBSCAN are determined through a K-distance graph, all space-time samples in a clustering training set are clustered, and a power utilization mode cluster with continuous time sequence is output; the second dense clustering is performed by using a DBSCAN algorithm, taking the mode cluster generated by the first refocus as a unit, performing DBSCAN secondary clustering on each time sequence mode cluster according to the power consumption mode similarity and the price sensitivity coefficient similarity as double measurement standards on samples in each cluster, and outputting subdivision mode clusters; and the third heavy business clustering is used for checking whether samples in the second sub-division mode cluster are in accordance with business constraints of corresponding industries one by one, and eliminating cross-industry error clustering samples which are not in accordance with the constraints; and finally, merging sub-clusters with the similarity more than or equal to 0.9 according to the service constraint checking result, and finally determining 10-15 typical power consumption mode clusters.
  4. 4. The multi-dimensional attribute power consumer time-sharing power cluster analysis prediction method according to claim 3, wherein the multi-branch collaborative prediction model comprises: A transducer encoder with time position codes is used as a time sequence branch, a 24-hour electric quantity weight sequence for historic 7 days of a user to be predicted and an hour-by-hour electric power spot price sequence for historic 7 days are input, and a time sequence feature vector is calculated; Adopting LightGBM model as attribute branch, inputting user basic attribute feature, market attribute feature and target cluster information, and calculating to obtain attribute feature vector; Adopting a 3-layer MLP network as a business constraint branch, inputting industry production specification features and business feature vectors, and obtaining business feature vectors and business compliance pre-scores; And (3) carrying out weighted fusion on the three branch outputs through an attention mechanism, generating fusion feature vectors to access the full-connection layer, outputting the electric quantity weight of 1 hour by each sub-layer, and finally obtaining a 24-dimensional time-sharing electric quantity weight prediction result.
  5. 5. The method for predicting the time-sharing electric quantity cluster analysis of the multi-dimensional electric power user according to claim 4, wherein the method for dynamically adjusting the electric quantity weight at the time of the initial 24 hours by combining the electric power spot market price signal and the price sensitivity level of the user to be predicted and synchronously executing the business rule verification and correction comprises the following steps: Carrying out period clustering on a day-ahead price curve of a day to be predicted by adopting a K-means clustering algorithm, wherein a clustering result comprises a price peak value section, a price flat value section and a price valley value section; and dynamically adjusting and normalizing the initial 24-hour time-sharing electric quantity weight according to the user sensitivity level and the price fluctuation rate, traversing the normalized weight according to the industry business rule and carrying out weight correction on the violation item according to the minimum adjustment principle until the final 24-hour time-sharing electric quantity weight is output without the violation item.
  6. 6. The method for predicting the time-sharing electric quantity cluster analysis of the multi-dimensional electric power consumer according to claim 5, wherein the dynamically adjusting the initial 24-hour time-sharing electric quantity weight according to the user sensitivity level and the price fluctuation rate comprises the following steps: The high-sensitivity user adjusts, aiming at three sections of price peak value, flat value and valley value, the price fluctuation rate and the sensitivity coefficient are combined, and the electric quantity weight in the initial 24 hours is dynamically adjusted through multiplication factors; The middle sensitive user adjusts, only carries on the weight adjustment to the peak value section and valley value section of the price, the flat value section maintains the initial weight; And low sensitivity user adjustment, only when the price peak section fluctuation rate exceeds 0.2, the non-core load weight is adjusted downwards.
  7. 7. A prediction system according to any one of claims 1-6, wherein the prediction system comprises: the data acquisition unit is used for acquiring basic data, electric power spot market data and business data of a user to be predicted; the data preprocessing unit is used for extracting multidimensional prediction input characteristics comprising time sequence characteristics, market characteristics and business characteristics from the acquired data; The first processing unit calls a pre-trained typical power consumption mode clustering template, calculates the similarity between a user to be predicted and each clustering template based on multidimensional prediction input characteristics, and determines target clusters and corresponding typical electric quantity curves; The second processing unit inputs the multidimensional prediction input characteristics and the target clustering information into the multi-branch collaborative prediction model to predict, and an initial 24-hour time-sharing electric quantity weight of a user to be predicted is obtained; The third processing unit is used for dynamically adjusting the power weight of the initial 24-hour time-sharing power by combining the power spot market price signal and the price sensitivity level of the user to be predicted, and synchronously executing the check and correction of the business rule of the industry; and the output unit outputs the optimized time-sharing electric quantity weight according to the daily, daily and real-time multiple time scales and synchronously generates a layered interpretation report containing characteristic contribution and service compliance.
  8. 8. The prediction system of claim 7 wherein the first processing unit further comprises a typical power usage pattern cluster template comprising: the data set module is used for acquiring multi-source training data of 3 years in history and dividing the multi-source training data into a clustering training set and a verification set according to a proportion; The clustering module is used for carrying out triple clustering training on the training set, wherein the triple clustering comprises a first heavy space-time clustering, a second heavy density clustering and a third heavy business clustering; and the output module takes the median sequence of each cluster as a typical electric quantity curve template of the cluster.
  9. 9. The prediction system of claim 8, wherein the second processing unit further comprises a multi-branch co-prediction model comprising: The input layer is used for converging the full-quantity features required by the model and classifying and integrating the multi-source data according to the branch requirements; The parallel branch layer comprises a time sequence branch, an attribute branch and a business constraint branch, wherein a transducer encoder with time position coding is adopted as the time sequence branch, a 24-hour electric quantity weight sequence for 7 days of a user to be predicted and an hour-by-hour electric power spot price sequence for 7 days of the history are input, a time sequence feature vector is calculated, a LightGBM model is adopted as the attribute branch, a user basic attribute feature, market attribute feature and target cluster information are input, and an attribute feature vector is calculated; The fusion layer is used for carrying out weighted fusion on the three branch outputs through an attention mechanism, and generating fusion feature vectors to be connected into full connection for output; and the output layer is used for outputting the electric quantity weight of 1 hour for each sub-layer, and finally obtaining a 24-dimensional time-sharing electric quantity weight prediction result.
  10. 10. A multi-dimensional attribute power consumer time-sharing electric quantity cluster analysis prediction model, which is a typical electricity consumption mode cluster template or a multi-branch collaborative prediction model in the multi-dimensional attribute power consumer time-sharing electric quantity cluster analysis prediction method according to any one of claims 1-6.

Description

Multi-dimensional attribute power consumer time-sharing electric quantity clustering analysis prediction method and system Technical Field The invention belongs to the technical field of power management, and particularly relates to a multi-dimensional attribute power consumer time-sharing electric quantity cluster analysis prediction method and system. Background With the continuous deepening of the power market reform in China, the power spot market is gradually established and perfected as a core platform for resource optimization configuration, the traditional power purchase and sale mode is broken, and the market pattern of competitive price surfing at the power generation side and autonomous selection at the user side is formed. In the power spot market environment, the power consumption behavior of a user is comprehensively influenced by multiple factors, and a time-sharing electric quantity curve presents remarkable complexity and diversity. The change of weather conditions such as temperature, humidity and the like directly affects the electricity consumption requirements of loads such as air conditioners, heating and the like from the external environment, the real-time price fluctuation of the electric power spot market can guide a user to adjust the electricity consumption period from the market level, particularly the electricity consumption behavior of commercial users and adjustable industrial loads is highly sensitive to price signals, and obvious differences can be generated in electricity consumption modes from the self attribute of the user in different industries such as continuous production industry, intermittent operation commercial retail, flexible electricity consumption resident users, different geographic positions and different distributed energy installation configurations (such as photovoltaic installation or energy storage installation or not). In addition, business factors such as production plan adjustment, equipment overhaul, holiday arrangement and the like of users can cause non-negligible influence on the time-sharing electric quantity curve, and the prediction difficulty is further increased. However, the existing power load prediction method is difficult to meet the high-precision reporting requirement of the power spot market, and mainly has the following two core defects: On the one hand, the characteristic dimension is single, and the key factors affecting the electricity utilization behavior cannot be covered comprehensively. The traditional load prediction method mostly takes historical electricity data and basic weather information as core input, focuses on time relevance and weather sensitivity of electricity consumption behaviors only, and ignores essential influence of multidimensional attributes such as user industry characteristics, geographic positions, distributed energy installation and the like on electricity consumption modes. Meanwhile, the conventional method generally does not include key information such as power spot market price signals, user service plans and the like, so that a prediction result is difficult to adapt to the adjustment of the power consumption behavior of a user in a marketization environment, and a marketization declaration decision of an electricity selling company cannot be supported. On the other hand, the user heterogeneity is not handled enough, and an effective classification and pattern recognition mechanism is lacking. The electricity consumption modes of users in different industries, different cities and different installation types are different, the electricity consumption curves of industrial continuous production users are stable, the night load is maintained at a higher level, the electricity consumption curves of commercial users show unimodal characteristics of daytime peaks and nighttime valleys and are obviously influenced by holidays, and the electricity consumption curves of resident users show bimodal forms of early peaks and late peaks. Most of the existing methods adopt a unified prediction model to predict all users in batches, and the classification processing is not carried out on the electricity utilization characteristics of different user groups, so that the model cannot be accurately adapted to the electricity utilization rules of various users. In part of the methods, although a simple clustering algorithm is adopted for user classification, the problems of fixed clustering parameters, unaccounted time sequence continuity, business constraint and the like exist, a typical electricity utilization mode with practical guiding significance is difficult to form, finally, the prediction precision is low, and the precision requirement of the daily declaration of an electricity selling company cannot be met. In summary, under the background of continuous and deep reform of the power market, the defects of the existing power load prediction method in feature coverage and user heterogeneity processing are i