An industry-based power load behavior recognition method based on deep learning

Abstract

This invention belongs to the field of power load identification technology, specifically a method for identifying industry-specific power load behavior based on deep learning. The specific steps of this method are as follows: hierarchical collection of multi-source heterogeneous data, construction of a standardized original dataset, construction of a deep learning-specific load sample set, building a multi-branch fusion deep learning load feature extraction model, supervised training and convergence optimization of the model, automatic extraction of full-dimensional load features, intelligent identification index output of industry-specific electricity consumption behavior, and load forecasting and demand-side management. This invention generates industry load sample sets by acquiring historical load data from industrial, commercial, public service, and residential users, combined with enterprise type, production shift, equipment operating status, and external environmental information. The industry load sample sets are then input into a deep learning model to extract abrupt changes, cycles, ramp-up, and sustained high load features from the load curves, thereby identifying industry-specific electricity consumption behavior patterns.

Description

Technical Field

[0001] This invention relates to the field of power load identification technology, specifically to a method for identifying power load behavior by industry based on deep learning. Background Technology

[0002] Accurate load sensing and refined electricity consumption behavior identification are the core foundational support for smart grid dispatching, demand response management, distribution network safety operation and maintenance, and efficient time-of-use pricing. Different industries have vastly different production schedules, equipment types, and production capacity conditions. The load curves, fluctuation patterns, and electricity consumption characteristics of industries such as industrial manufacturing, commercial supermarkets, residential communities, agricultural irrigation, and public buildings are drastically different. Accurately distinguishing industry categories and identifying the electricity consumption behavior patterns within each industry enables the power grid to achieve differentiated load regulation, precise fault location, investigation of electricity theft and anomalies, matching of renewable energy consumption, and dual-control statistical accounting of industry energy consumption. This has crucial practical significance for the stable and economical operation of the new power system.

[0003] However, current methods for identifying power load behavior still have the following shortcomings:

[0004] Existing power load identification methods mostly rely on single load time-series data, lacking multi-dimensional auxiliary information such as industry attributes, production conditions, and environment. They cannot distinguish the differentiated electricity consumption patterns of four types of users: industrial, commercial, public service, and residential. Conventional statistical algorithms are unable to autonomously extract the latent characteristics of load curves, such as periodicity, ramp-up, abrupt changes, and sustained high load. The classification of industry electricity consumption behavior is coarse, and the accuracy of abnormal load identification is low. This directly results in large load forecasting deviations and difficulties in implementing refined demand-side management and inspection.

[0005] To address this issue, a deep learning-based method for identifying industry-specific power load behavior is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based method for identifying industry-specific power load behavior, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying industry-specific power load behavior based on deep learning, the specific steps of which are as follows:

[0008] Step 1: Layered collection of multi-source heterogeneous data: Collect user data from four major categories: industrial, commercial, public service, and residential. User data includes electricity consumption data, industry attribute data, equipment operating condition data, and external environment data.

[0009] Step 2: Construct a standardized raw dataset: Remove data such as vacuum values ​​caused by disconnection, meter failure, and abnormal data due to exceeding the range; fill in missing load points using linear interpolation and the average value of the same period in the same industry; normalize continuous quantities using Min-Max to unify the units, and use one-hot encoding for discrete quantization of industry type and production shift; initially classify the samples according to four major categories: industrial, commercial, public service, and residential.

[0010] Step 3: Construct a dedicated deep learning workload sample set: Using the user's industry as the basic label, bind the corresponding production shift, equipment configuration, and environmental parameters as auxiliary features, divide the sample set into training set, validation set, and test set, complete the hierarchical encapsulation of the sample set, and adapt it to the input format of the deep learning model.

[0011] Step 4: Build a multi-branch fusion deep learning load feature extraction model: The backbone adopts a hybrid deep learning network of CNN and LSTM, and the model input is simultaneously connected to industry, equipment and environmental auxiliary features to achieve multi-source information fusion training;

[0012] Step 5: Supervised Model Training and Convergence Optimization: Supervised training is performed using a sample set labeled with industry and electricity consumption behavior tags. The network weights are jointly optimized using load behavior classification loss and anomaly recognition loss. The validation set is used to iteratively tune parameters to prevent overfitting. After convergence, the optimal recognition model is saved.

[0013] Step 6: Automatic extraction of full-dimensional load features: Input the load samples of the users to be identified into the trained deep learning model, and the network autonomously separates four types of core features from the load curve;

[0014] Step 7: Output of intelligent identification indicators for electricity consumption behavior by industry: Based on the extracted features, complete the behavior analysis and output three types of results in batches;

[0015] Step 8, Load Forecasting and Demand-Side Management: Integrate the identified industry electricity consumption patterns and anomaly tags into the power grid business system.

[0016] Preferably, in step one, the user side includes historical hourly and daily electricity load time-series data; the industry attribute data includes enterprise registration industry classification, production capacity, production process, and work schedule; the equipment operating condition data includes production equipment start-up and shutdown ledgers, high-power equipment switching records, and air conditioning and HVAC installed capacity; and the external environment data includes daily temperature, humidity, holidays, workdays, rest days, and seasonal variation data.

[0017] Preferably, the CNN branch model is used to extract short-term local features of the load, including load sudden spikes, short-term ramp-ups, and instantaneous start-stop jump features; the LSTM branch model is used to mine long-term periodic patterns, including daily cycles, weekly cycles, seasonal electricity consumption patterns, and long-term continuous high load features.

[0018] Preferably, in step six, the core features include abrupt change features, periodic features, ramp-up features, and sustained high load features; abrupt change features include sudden increases and decreases in load caused by equipment start-up and shutdown, temporary overtime, and fault shutdown; periodic features include periodic electricity consumption patterns formed by fixed production shifts, weekdays, and weekends; ramp-up features include a slow change curve of gradually increasing load when starting up and gradually decreasing load when stopping production; sustained high load features include steady-state electricity consumption characteristics of continuous full-load production for 24 hours and multiple days.

[0019] Preferably, in step seven, the three types of results include industry load attribution category, typical energy consumption period, and abnormal load risk; industry load attribution category includes accurately determining the user's sub-sector of industry, commerce, public services, and residential; typical energy consumption period includes peak electricity consumption intervals, off-peak rest periods, and off-peak production periods for each industry; abnormal load risk includes abnormal power consumption exceeding production capacity, sudden increase in load due to electricity theft, and abnormal power consumption under no-load conditions due to equipment failure.

[0020] Preferably, the detailed steps of step eight are as follows: refine the regional load forecasting model according to industry, improve the accuracy of short-term and medium-to-long-term load forecasts, carry out targeted electricity consumption inspections and illegal electricity consumption supervision based on the abnormal load list, formulate orderly electricity consumption, demand response, and staggered production plans according to the typical energy consumption periods of each industry, and optimize demand-side resource management.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] This invention generates an industry load sample set by acquiring historical load data from industrial, commercial, public service, and residential users, and combining this data with enterprise type, production shift, equipment operating status, and external environmental information. The industry load sample set is then input into a deep learning model to extract abrupt changes, cycles, ramp-up characteristics, and sustained high load features from the load curves, identifying industry electricity consumption behavior patterns. Based on the identification results, the invention outputs industry load categories, typical energy consumption periods, and abnormal load risks, enabling refined analysis of load characteristics in different industries and improving the accuracy of load forecasting, electricity consumption supervision, and demand-side management.

[0023] (1) This invention overcomes the technical problems of traditional load identification data being single-dimensional, of poor quality, and lacking industry-specific sample systems through steps one, two, and three. It no longer relies solely on load time-series data, but collects electricity consumption time-series data, industry attribute data, equipment operating condition data, and external environment data from four major categories of users: industrial, commercial, public service, and residential. This covers multi-dimensional related information such as production capacity, production schedule, equipment start-up and shutdown, temperature and humidity, and holidays and seasons, completely solving the defects of traditional models that rely only on a few electrical features and ignore industry production patterns and environmental interference factors. At the same time, through the removal of abnormal bad data, interpolation and completion of missing data, data normalization and discrete feature encoding, and industry label classification and encapsulation, the standardization processing of multi-source heterogeneous data is completed, and an industry-specific load sample set adapted for deep learning training is constructed. This solves the technical problems of traditional datasets being mixed and disordered, with serious coupling of industry features, and unable to support refined modeling, laying a high-quality data foundation for subsequent accurate identification of industry electricity consumption behavior.

[0024] (2) This invention addresses the core technical bottlenecks of traditional shallow machine learning models by building a multi-branch fusion deep learning model of CNN and LSTM in step four. This model has weak feature extraction capabilities, cannot take into account both short-term transient features and long-term patterns, and has poor industry adaptability. The CNN branch is specifically designed to mine short-term local load features and accurately capture transient power consumption changes such as equipment start-up and shutdown, load peaks, short-term ramp-ups, and instantaneous jumps. The LSTM branch deeply mines long-term periodic power consumption patterns and continuous high-load steady-state features overcoming the shortcomings of traditional models that can only identify simple linear patterns and cannot decompose multi-scale coupled load features. At the same time, the model input integrates industry attributes, equipment operating conditions, and external environmental auxiliary features to achieve deep fusion training of multi-source information. This solves the problems of low recognition of load features of different industries, dilution of features of small and medium-sized industries, and high industry misjudgment rate of traditional unified models.

[0025] (3) This invention solves the problems of overfitting, poor robustness, and weak generalization ability of traditional models through the joint loss supervised training and iterative optimization mechanism in step five. It adopts a joint optimization strategy of industry classification loss and anomaly recognition loss, relies on a well-labeled industry sample set for supervised training, and combines iterative parameter tuning on the validation set to suppress model overfitting, retain the optimal converged model, and greatly improve the recognition stability and cross-scene generalization ability of the model under complex working conditions and noisy data, avoiding the problem of a significant decrease in recognition accuracy caused by data noise and working condition fluctuations in traditional models;

[0026] (4) This invention extracts four core load characteristics—mutation characteristics, periodic characteristics, ramp-up characteristics, and continuous high-load characteristics—in step six, accurately corresponding to the production electricity consumption behavior patterns of different industries, thus solving the defect of traditional technology in being unable to finely decompose industry electricity consumption behavior. By distinguishing the mutation loads formed by equipment start-up and shutdown, temporary overtime, and fault shutdown, the periodic loads formed by production shifts and work schedules, the ramp-up loads formed by equipment start-up and shutdown load increases and decreases, and the steady-state high-load loads of continuous full production, it achieves accurate decomposition of the differentiated energy consumption patterns of various industries such as industry, commerce, public services, and residential, breaking through the technical shortcomings of traditional technology, which can only roughly identify the major load categories and cannot subdivide production conditions and electricity consumption status.

[0027] (5) This invention solves the problems of traditional load identification lacking refined behavior analysis and abnormal risk screening by batch outputting three types of identification results in step seven: industry load classification, typical energy consumption period, and abnormal load risk. It can accurately distinguish the user's subdivided industry classification, locate the peak and valley electricity consumption period and off-peak production pattern of each industry, and intelligently mark abnormal loads such as overcapacity power consumption, sudden increase in electricity theft, and equipment no-load failure. It realizes the upgrade from single load identification to a full-dimensional refined analysis of industry classification, behavior analysis, and risk warning, making up for the defects of traditional electricity supervision that rely on manual thresholds, have low accuracy, and poor targeting.

[0028] (6) This invention deeply integrates the industry-specific load identification results with the actual power grid operations through step eight, solving the technical problems of the disconnect between traditional load identification and power grid dispatching, demand-side management, electricity consumption supervision, and low load forecast accuracy. Based on the real electricity consumption patterns of various industries, the regional load forecast model is corrected, which greatly improves the accuracy of long-term and short-term load forecasts; based on the abnormal load list, targeted electricity consumption inspection and violation supervision are realized; at the same time, orderly electricity consumption, staggered production, and demand response strategies are formulated according to the differentiated energy consumption periods of various industries, realizing industry-specific, refined, and intelligent demand-side resource management, and completely solving the industry pain points of traditional technology being unable to adapt to the differentiated power grid operations of multiple industries, having extensive management and control, and having low prediction accuracy. Attached Figure Description

[0029] Figure 1 A flowchart illustrating the steps of a power load behavior identification method. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example:

[0032] Please see Figure 1 The present invention provides a technical solution:

[0033] A deep learning-based method for identifying industry-specific power load behavior, the specific steps of which are as follows:

[0034] Step 1: Layered Collection of Multi-Source Heterogeneous Data: Data is collected from four major categories of users: industrial, commercial, public service, and residential. This user data includes electricity consumption data, industry attribute data, equipment operating condition data, and external environment data. This approach overcomes the limitations of single-load time-series data, horizontally covering four categories of information: electricity consumption time-series, industry attributes, equipment operating conditions, and external environment. This categorizes users into four major user systems: industrial, commercial, public service, and residential. It provides a complete data source for subsequently distinguishing differentiated electricity consumption patterns across different industries and eliminating environmental and equipment interference factors, thus solving the problem of identification confusion caused by relying solely on electricity values ​​and ignoring production and external influencing factors.

[0035] Step 2: Construct a standardized raw dataset: Remove data containing missing values ​​due to disconnection, meter malfunction, or exceeding the specified range; fill vacant load points using linear interpolation and the industry average for the same period; unify the dimensions of continuous quantities using Min-Max normalization, and perform one-hot encoding for industry type and production shift, initially classifying samples according to four major categories: industrial, commercial, public service, and residential; remove null values ​​and exceeding the specified range, and filter out invalid and distorted data caused by meter malfunction or communication disconnection; use a dual strategy to fill vacant locations: linear interpolation adapts to short-term minor gaps, while the industry average for the same period fills long-term gaps to match industry electricity consumption patterns, ensuring data continuity; Min-Max normalization eliminates the differences in dimensions of different physical quantities such as power, temperature, and humidity, and one-hot encoding converts the industry and shift categories into model-calcifiable values; pre-label and classify by the four major industries to achieve data hierarchical organization and reduce feature coupling interference during subsequent model training.

[0036] Step 3: Construct a dedicated deep learning workload sample set: Using the user's industry as the basic label, bind corresponding production shifts, equipment configurations, and environmental parameters as auxiliary features, divide the dataset into training, validation, and test sets, and complete the hierarchical encapsulation of the sample set to adapt to the input format of the deep learning model; bind samples with industry as the main label and shifts, equipment, and environment as auxiliary features to establish a standardized sample structure with a one-to-one correspondence between labels and features; divide the dataset into three groups: training, validation, and test, to realize a three-stage process of training fitting, mid-process parameter tuning and verification, and independent evaluation of the final effect; encapsulate the samples according to the network input format to adapt to the input specifications of the CNN-LSTM hybrid model, solving the problem that messy raw data cannot be directly fed into the deep learning network for training;

[0037] Step 4: Build a multi-branch fusion deep learning load feature extraction model: The backbone adopts a hybrid deep learning network of CNN and LSTM, and the model input simultaneously accesses auxiliary features from the industry, equipment, and environment to achieve multi-source information fusion training; the dual backbone branches extract load features at different scales, making up for the deficiency that a single network can only capture a single time scale; the simultaneous access of auxiliary features from the industry, equipment, and environment enables multi-source fusion training of power time series data and business static features; compared with a globally unified model, the fusion architecture can adapt to the huge differences in load fluctuation patterns of the four major industries, improving the industry-specific identification.

[0038] Step 5: Supervised Model Training and Convergence Optimization: Supervised training is performed using a sample set labeled with industry and electricity consumption behavior tags. The network weights are jointly optimized using load behavior classification loss and anomaly recognition loss. Iterative parameter tuning is used with the validation set to prevent overfitting. The optimal recognition model is saved after convergence. Supervised training with dual industry and electricity consumption behavior tags is performed to accurately learn the correspondence between industry categories and electricity consumption conditions. The weights are jointly optimized using classification loss and anomaly recognition loss to simultaneously consider industry discrimination accuracy and abnormal load recognition capability. Iterative parameter tuning is performed with the validation set to suppress overfitting, and the optimal converged model is selected to improve the model's noise resistance and cross-scenario generalization ability, avoiding on-site recognition failure due to training overfitting.

[0039] Step Six: Automatic Extraction of Full-Dimensional Load Features: Input the load samples of the users to be identified into the trained deep learning model. The network autonomously separates four types of core features from the load curve. The trained model autonomously analyzes the load curve and separates four types of core features: sudden change, periodicity, ramp-up, and continuous high load. No manual feature design and selection is required. Deep learning automatically uncovers deep and hidden electricity consumption patterns, overcoming the shortcomings of traditional manual feature extraction, which is one-sided and superficial.

[0040] Step 7: Intelligent Identification Index Output for Industry-Specific Electricity Consumption Behavior: Based on extracted features, behavior analysis is completed, and three types of results are output in batches; three types of analysis results are output at once: industry classification, typical peak and valley energy consumption periods, and abnormal load risk; upgrading from simple category identification to integrated analysis of industry classification, behavior period analysis, and fault electricity theft risk warning, supporting refined power grid management;

[0041] Step 8, Load Forecasting and Demand-Side Management: Integrate the identified industry electricity consumption patterns and anomaly tags into the power grid business system, and integrate the identified industry electricity consumption patterns into the business system to form a complete business closed loop of identification, forecasting, supervision, and control; calibrate the load forecasting model by industry to improve the accuracy of long-term and short-term load forecasts; accurately investigate electricity theft and equipment failure power consumption based on the anomaly list; and formulate orderly electricity consumption, demand response, and peak-shaving plans according to the peak and valley periods of each industry to achieve differentiated and refined demand-side management.

[0042] In step one, the user side includes historical hourly and daily electricity load time-series data; the industry attribute data includes enterprise registration industry classification, production capacity, production process, and work schedule; the equipment condition data includes production equipment start-up and shutdown logs, high-power equipment switching records, and HVAC installed capacity; the external environment data includes daily temperature, humidity, holidays, weekdays, rest days, and seasonal variation data. The electricity consumption side locks in hourly and daily time-series loads to ensure the time-series sequence length meets the requirements for long and short time-series feature mining. Industry attributes are linked to production capacity, process, and work schedule to explain the production root causes of industry load differences. Equipment condition data is associated with high-power equipment and HVAC installed capacity to explain the equipment-related causes of load abrupt changes and ramp-ups. The environment data includes temperature, humidity, holidays, and seasons to eliminate the interference of weather and work schedules on the load curve.

[0043] The CNN branch model is used to extract short-term local load features, including load spikes, short-term ramps, and instantaneous start-stop transitions. The LSTM branch model is used to mine long-term periodic patterns, including daily, weekly, seasonal electricity consumption patterns, and long-term sustained high load characteristics. The CNN is responsible for short-term local instantaneous changes: spikes, ramps, and start-stop transitions, capturing rapid fluctuations caused by instantaneous equipment switching. The LSTM is responsible for long-term periodic patterns: daily, weekly, seasonal cycles, and long-term full-load steady state.

[0044] In step six, the core features include abrupt change features, periodic features, ramp-up features, and sustained high load features. Abrupt change features include sudden increases and decreases in load caused by equipment start-up and shutdown, temporary overtime, and machine failure shutdowns. Periodic features include periodic electricity consumption patterns formed by fixed production shifts, weekdays, and weekends. Ramp-up features include a slow change curve of gradually increasing load during startup and gradually decreasing load during shutdown. Sustained high load features include steady-state electricity consumption characteristics of continuous 24-hour and multi-day uninterrupted full-load production. A one-to-one mapping relationship is established between the load curve shape and actual production electricity consumption behavior to ensure that the features extracted by the model accurately correspond to actual operating conditions, making the behavior interpretable and the identification results verifiable.

[0045] In step seven, the three types of results include industry load attribution category, typical energy consumption period, and abnormal load risk. Industry load attribution category includes accurately determining the user's sub-sector: industrial, commercial, public service, and residential. Typical energy consumption period includes peak electricity consumption intervals, off-peak rest periods, and staggered production periods for each industry. Abnormal load risk includes abnormal power consumption exceeding production capacity, sudden increases in load due to electricity theft, and abnormal power consumption under no-load conditions caused by equipment malfunctions. The identification output dimensions are limited to distinguish between normal industry behavior and abnormal malfunctions / violations, balancing the two major business needs of normal energy consumption analysis and safety inspection.

[0046] The detailed steps of step eight are as follows: Refine the regional load forecasting model according to industry sectors to improve the accuracy of short-term and medium-to-long-term load forecasts; conduct targeted electricity consumption inspections and illegal electricity consumption supervision based on abnormal load lists; formulate orderly electricity consumption, demand response, and staggered production plans according to typical energy consumption periods of each industry; and optimize demand-side resource management. Establish a pathway between the algorithm model and actual power grid operation, proving that the method has practical engineering value and is not limited to simple algorithm recognition, but can directly empower the entire business chain of dispatching, inspection, and demand-side operations.

[0047] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or basic characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description. Therefore, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based industry-specific power load behavior recognition method, characterized by, The specific steps of this method are as follows: Step 1: Layered collection of multi-source heterogeneous data: Collect user data from four major categories: industrial, commercial, public service, and residential. User data includes electricity consumption data, industry attribute data, equipment operating condition data, and external environment data. Step 2: Construct a standardized raw dataset: Remove data with missing vacuum values ​​caused by disconnection, meter failure, or abnormal data exceeding the range; fill in missing load points using linear interpolation and the average value of the same period in the same industry; normalize continuous quantities using Min-Max to unify the units, and use one-hot encoding for discrete quantization of industry type and production shift; initially classify the samples according to four major categories: industrial, commercial, public service, and residential. Step 3: Construct a dedicated deep learning workload sample set: Using the user's industry as the basic label, bind the corresponding production shift, equipment configuration, and environmental parameters as auxiliary features, divide the sample set into training set, validation set, and test set, complete the hierarchical encapsulation of the sample set, and adapt it to the input format of the deep learning model. Step 4: Build a multi-branch fusion deep learning load feature extraction model: The backbone adopts a hybrid deep learning network of CNN and LSTM, and the model input is simultaneously connected to industry, equipment and environmental auxiliary features to realize multi-source information fusion training; Step 5: Supervised Model Training and Convergence Optimization: Supervised training is performed using a sample set labeled with industry and electricity consumption behavior tags. The network weights are jointly optimized using load behavior classification loss and anomaly recognition loss. The validation set is used to iteratively tune parameters to prevent overfitting. After convergence, the optimal recognition model is saved. Step 6: Automatic extraction of full-dimensional load features: Input the load samples of the users to be identified into the trained deep learning model, and the network autonomously separates four types of core features from the load curve; Step 7: Output of intelligent identification indicators for electricity consumption behavior by industry: Based on the extracted features, complete the behavior analysis and output three types of results in batches; Step 8, Load Forecasting and Demand-Side Management: Integrate the identified industry electricity consumption patterns and anomaly tags into the power grid business system.

2. The industry-specific power load behavior recognition method based on deep learning according to claim 1, characterized in that: In step one, the user side includes historical hourly and daily electricity load time-series data; the industry attribute data includes enterprise registration industry classification, production capacity, production process, and work schedule; the equipment operating condition data includes production equipment start-up and shutdown ledgers, high-power equipment switching records, and air conditioning and HVAC installed capacity; the external environment data includes daily temperature, humidity, holidays, workdays, rest days, and seasonal variation data.

3. The industry-specific power load behavior recognition method based on deep learning according to claim 1, characterized in that: The CNN branch model is used to extract short-term local features of the load, including load spikes, short-term ramp-ups, and instantaneous start-stop jumps. The LSTM branch model is used to mine long-term periodic patterns, including daily, weekly, seasonal electricity consumption patterns, and long-term continuous high load features.

4. The industry-specific power load behavior recognition method based on deep learning according to claim 1, characterized in that: In step six, the core features include mutation features, periodic features, climbing features, and continuous high load features; Sudden changes include sudden increases and decreases in load caused by equipment start-up and shutdown, temporary overtime work, and machine failures; periodic changes include periodic electricity consumption patterns formed by fixed production shifts, weekdays, and weekends. The ramp-up characteristics include a slow change curve of gradually increasing load when starting up and gradually decreasing load when stopping production; the continuous high load characteristics include the steady-state power consumption characteristics of continuous full-load production for 24 hours and multiple days.

5. The industry-specific power load behavior recognition method based on deep learning according to claim 1, characterized in that: In step seven, the three types of results include industry load attribution category, typical energy consumption period, and abnormal load risk. Industry load attribution category includes accurately determining the user's sub-sectors of industry, commerce, public services, and residential. Typical energy consumption period includes peak electricity consumption intervals, off-peak rest periods, and staggered production periods for each industry. Abnormal load risk includes abnormal power consumption due to overcapacity, sudden increase in load due to electricity theft, and abnormal power consumption due to equipment failure under no-load conditions.

6. The industry-specific power load behavior recognition method based on deep learning according to claim 1, characterized in that: The detailed steps of step eight are as follows: refine the regional load forecasting model according to industry, improve the accuracy of short-term and medium-to-long-term load forecasts, carry out targeted electricity consumption inspections and supervision of illegal electricity consumption based on the abnormal load list, formulate orderly electricity consumption, demand response, and staggered production plans according to the typical energy consumption periods of each industry, and optimize demand-side resource management.

Images