Information processing method and system for electric power marketing management system

Abstract

The invention belongs to the technical field of power marketing, and particularly discloses an information processing method and system for a power marketing management system. The method comprises the following steps: S1, data acquisition: acquiring electric power data through a sensor; s2, preprocessing data, calculating an average value and a standard deviation, removing abnormal values and filling missing values; s3, feature extraction: calculating a power consumption feature value and a load fluctuation feature value, and dynamically adjusting the acquisition frequency; s4, load prediction: predicting a power load value by using a neural network model; and S5, marketing decision making: making an electricity price, resource allocation and promotion activity strategy according to the predicted value. The system comprises a data acquisition module, a preprocessing module, a feature extraction module, a load prediction module, a marketing decision module, a data storage module and a communication module. The problems that a traditional power load prediction method is low in precision and an existing marketing decision lags behind and is difficult to meet individual requirements are solved. Experiments prove that the method has a good application effect in different scenes, and the efficiency and accuracy of power marketing management are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power marketing, and in particular, relates to an information processing method and system for a power marketing management system. Background Art

[0002] Power marketing management occupies a key position in the operation of modern power systems. It is an important means for power companies to achieve optimal resource allocation, improve economic benefits and meet social electricity demand. Accurate power load forecasting is of great significance for formulating reasonable plans and scheduling plans, improving equipment utilization, and ensuring safe and stable operation of the system. This is a key task in power marketing management and is directly related to the efficiency of power supply management. In recent years, electricity consumption has increased significantly, and the difficulty of power load forecasting has increased accordingly. Traditional load forecasting methods face many challenges. Traditional methods such as time series analysis are difficult to cope with nonlinear and non-stationary changes in power load. In practice, factors such as seasonal changes in industrial production, significant differences in residents' electricity consumption behavior on weekdays and weekends, and extreme weather conditions make the load change pattern extremely complex. Traditional methods are difficult to accurately capture, resulting in poor prediction accuracy and unable to meet the strict requirements of power marketing management for load forecasting accuracy, which in turn affects the effective development of power energy-saving project research and development, because accurate load forecasting is an important basis for implementing power energy-saving measures, and power energy-saving project research and development is also an important part of optimizing resource utilization in power supply management.

[0003] The existing power marketing decision-making methods have obvious lags in dealing with the real-time changing power market environment. Taking the patent document CN118981557A as an example, the patent discloses a visualization method and system for power storage allocation and scheduling, which mainly relies on historical data for decision-making and lacks effective integration and utilization of real-time data. In today's rapidly changing power market, power demand changes rapidly. Relying only on historical data to formulate marketing strategies cannot respond to the dynamic changes of the market in a timely manner, which also has an adverse impact on the research and development of power energy-saving projects, and it is difficult to adjust energy-saving strategies according to real-time conditions. For example, when sudden high temperature weather causes a sharp increase in power demand, traditional decision-making methods are difficult to quickly adjust resource allocation and electricity price strategies, thereby affecting the stability of power supply and the economic benefits of enterprises, and are also not conducive to guiding users to use electricity reasonably to achieve energy-saving goals. In addition, the existing decision-making methods are also insufficient in dealing with the personalized needs of different regions and different user groups, making it difficult to achieve precision marketing and promoting power energy-saving measures in different scenarios. Precision marketing and personalized services are key factors in modern power marketing management in improving competitiveness and service quality.

[0004] Data quality is a key link in power marketing management, which directly affects the accuracy and effectiveness of various management activities. In the process of power data collection, due to factors such as sensor aging and transmission interference, data is often plagued by outliers and missing values. Outliers may be caused by sensor failure or instantaneous power interference. If they are not identified and processed, they will greatly mislead the subsequent load forecasting model, causing the model to deviate from the true power load trend, resulting in the inability to accurately grasp the power consumption in the development of power energy-saving projects, and it is difficult to formulate effective energy-saving plans. Missing values ​​will destroy the integrity of the data, affect the accuracy of data analysis, and then lead to deviations in the formulation of marketing strategies, which is also not conducive to the accurate formulation of energy-saving strategies in the development of power energy-saving projects. For example, during peak load periods, if the power data of key nodes is missing, power companies may misjudge power demand and fail to allocate resources in a timely manner, affecting power supply reliability and customer satisfaction, and not conducive to the reasonable implementation of energy-saving measures.

[0005] The application of neural network technology has become a research hotspot. Therefore, how to use the prediction results of the neural network model can provide strong support for decision-making in power marketing management, including the formulation of energy-saving strategies in the research and development of power energy-saving projects, such as reasonably adjusting electricity prices to guide users to use electricity at off-peak times, achieve power load balance, and achieve energy-saving goals; accurate resource allocation ensures efficient and reasonable power supply and avoids energy waste; effective promotional activities stimulate users to adopt energy-saving equipment and energy-saving electricity methods, improve corporate profits and service quality, enhance corporate market competitiveness, and promote the development of power energy-saving projects. Summary of the invention

[0006] The present invention aims to provide an information processing method and system for a power marketing management system, which overcomes the defects of the prior art and improves the efficiency and accuracy of power marketing management through precise data collection and preprocessing, effective feature extraction, advanced neural network model load forecasting and scientific marketing decision making.

[0007] To achieve the above object, the present invention adopts the following technical solution: an information processing method for a power marketing management system, which comprises the following steps: S1 Data Collection: Collect power data through sensors installed at various nodes of the power system and mark the collected data as D ={( V 1 , I 1 , PF 1 , t 1 ),( V 2 , I 2 , PF 2 , t 2 ),…,( V n , I n , PF n , t n )},in n is the number of collections, V is the voltage value, I is the current value, PF is the power factor, t These sensors can obtain various key parameters in the operation of the power system in real time and accurately, providing a rich data basis for subsequent analysis.

[0008] S2 data preprocessing: preprocess the data D collected by S1. First, calculate the average value of voltage, current and power factor. V av e 、I av e 、PF av e and standard deviation s .

[0009] The standard deviation σ is calculated as ,in x i For each data value of voltage, current, and power factor in the data set, is the average value of all data values ​​in the data set. Then, according to the preset threshold k 1 、k 2 、k 3 Determine abnormal value, if the voltage value V satisfy |VV av e |>k 1 s , or current value I satisfy |I- av e |>k 2 s , or power factor PF satisfy |PF-PF av e|>k 3 s , it is determined as an outlier and removed. For missing values, the average value of adjacent data is used to fill in the preprocessed data. D' .

[0010] S3 feature extraction: based on the data preprocessed by S2 D' , accurately calculate the power consumption characteristic value E , the calculation method is ,in V i For the i The voltage value is collected. I i For the i The current value is collected. PF i is the power factor collected for the i-th time, t i is the i-th acquisition timestamp, Δt The data is collected once every scheduled time interval, i.e. the collection frequency. The collection frequency is dynamically adjusted according to the load fluctuation characteristics of the power system. Δt The adjustment method is: , in T is the initial set collection time interval, α、β To adjust the parameters, P is the total power of the power system currently monitored, P 0 is the preset power threshold. In addition, the power load fluctuation characteristic value is also calculated F , the calculation method is: , in n Represents the total number of times data is collected. i Indicates the sequential index of data collection, E i For the i The power consumption characteristic value is calculated based on the collected data.

[0011] S4 Load Forecasting: Input the power consumption characteristic value E into a pre-trained load forecasting model, wherein the load forecasting model adopts any one of the neural network models of MFNN, LSTM, and CNN, and is preferably an MFNN model. The number of neurons in the input layer of the MFNN model is determined according to the dimension of the power consumption characteristic value, and the hidden layer includes at least one layer. The number of neurons in the hidden layer h is determined according to the following formula: , in mis the number of neurons in the input layer, l is the number of neurons in the output layer, k To adjust the coefficient, the output layer outputs the predicted power load value L for the future time period.

[0012] The training process of the neural network model is as follows: S4.1 collect historical power data, process it according to the above data collection and preprocessing steps, and obtain historical power consumption characteristic values ​​and corresponding historical actual power load values; S4.2 takes the historical power consumption characteristic value as input and the historical actual power load value as output, and uses the back propagation algorithm to adjust the weights and biases of the neural network model for training to minimize the loss function value. The mean square error MSE is used as the loss function during the training process. The loss function calculation formula is: , in q is the number of training samples, L j For the j The actual power load value of the samples, For the j The predicted power load value of samples; S4.3 continuously repeats the training process of S4.2 until the loss function value converges to between 0.1 and 1, which means that the neural network model meets the required prediction ability.

[0013] S5 Marketing decision: Formulate power marketing strategies based on the predicted power load value L, including adjusting electricity price strategies, power resource allocation strategies, and power promotion activity strategies.

[0014] The electricity price adjustment strategy includes: When the predicted power load value is greater than the preset high load threshold, the peak period electricity price is increased according to the following formula: , in r h To increase the amplitude, c is the electricity price adjustment coefficient, L To predict the power load value, L h To preset a high load threshold; When the predicted power load value is less than the preset low load threshold, the off-peak period electricity price is reduced according to the following formula: , in r l To reduce the amplitude, d is the electricity price adjustment coefficient, LTo predict the power load value, L l The preset low load threshold.

[0015] The power resource allocation strategy includes: According to the predicted power load values ​​of different regions, power resources are preferentially allocated to areas with higher load values ​​and obvious growth trends, calculated according to the following formula: , in p For the distribution ratio, L To predict the power load value, L i is the predicted power load value of the ith region, m is the number of regions, i Assign adjustment factors to resources.

[0016] The electricity promotion activity strategies include: When the predicted power load value is within the preset promotion load range, the power package discount information is pushed to the user, calculated according to the following formula: , Where d is the preferential strength, e is the preferential intensity adjustment coefficient, L To predict the power load value, [L p1 ,L p2 ] is the preset promotion load range.

[0017] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention utilizes a data acquisition module to access different types of sensors such as voltage sensors, current sensors, and power factor sensors to obtain real-time and accurate data, and then uses a data preprocessing module to remove abnormal values ​​from the collected data and properly fill in missing values. After this processing, the data is accurate and complete, laying a reliable foundation for subsequent load forecasting model training and marketing decision-making, effectively avoiding decision-making deviations and risks caused by data defects, and effectively ensuring the scientificity and accuracy of decision-making.

[0018] The present invention uses a feature extraction module to flexibly adjust the acquisition frequency according to the system power dynamics, deeply mine the data content, accurately extract the characteristic values ​​of power consumption and load fluctuations, and closely fit the actual operation rhythm of the power system. The load forecasting module adopts an advanced neural network model, whose powerful nonlinear fitting ability can fully analyze complex load data, accurately capture the context of load changes, and the prediction accuracy far exceeds that of traditional methods. It provides solid data support for marketing decisions, greatly reduces the operational risks caused by prediction errors, and helps enterprises accurately predict load trends and scientifically plan resource allocation.

[0019] The present invention adopts the present invention to customize multiple marketing strategies guided by accurate load forecasting, with remarkable results. The electricity price adjustment strategy is flexible and adaptable according to the load threshold, raising prices during peak hours and reducing prices during valley hours, accurately guiding users to reduce peak hours and fill valley hours, use electricity scientifically, effectively balance the contradiction between electricity supply and demand, improve equipment utilization, and reduce energy loss; the resource allocation strategy accurately allocates electricity resources according to the regional load forecast value, giving priority to guaranteeing the electricity demand of key areas and high-growth areas, ensuring stable and orderly electricity consumption for production and life, and smooth and unimpeded operation; the promotion activity strategy accurately captures the promotion opportunities, pushes preferential packages in a timely manner, strongly stimulates users' enthusiasm for electricity consumption, increases electricity sales, expands market share, and enhances corporate profitability and market influence.

[0020] The present invention has been rigorously verified by experiments. It utilizes the prediction results of the neural network model and selects the advantages of the MFNN model in the field of load forecasting. It meets the real-time needs of power marketing management and can quickly respond to system changes and instantly update the prediction model to provide timely and accurate basis for marketing decisions. The model architecture is simple and intuitive, deeply integrated with business scenarios, reducing application maintenance costs and technical barriers, and comprehensively improving the intelligence and efficiency of power marketing management, promoting the steady improvement of corporate operating efficiency and market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a flow chart of the information processing method of the power marketing management system in the present invention; Figure 2 This is a schematic diagram of the power marketing management system of the present invention; Figure 3 Comparison chart of loss values ​​and accuracy of the three models of MFNN, LSTM and CNN on the validation set and test set before and after data preprocessing in the present invention; Figure 4 This is a test result diagram of the data acquisition module in the present invention; Figure 5 This is a test result diagram of the data preprocessing module in the present invention; Figure 6 This is a test result diagram of the feature extraction module in the present invention; Figure 7 This is a test result diagram of the load prediction module in the present invention; Figure 8 This is a test result diagram of the marketing decision module in the present invention; Figure 9 This is a test result diagram of the power marketing management system integrated in the present invention; Figure 10 This is a bar chart comparing the prediction accuracy of the three models of MFNN, LSTM and CNN in the present invention; Figure 11This is a bar chart comparing the training time of the three models of MFNN, LSTM and CNN in the present invention. DETAILED DESCRIPTION

[0022] In order to further illustrate the technical solution of the present invention, the following is a combination of the entire project development process, and the relevant experimental data are selected. Figure 1 to 11 The present invention will be further described.

[0023] The present invention aims to provide an information processing method and system for a power marketing management system, and demonstrates its application and effects in different scenarios through multiple embodiments, so as to solve many problems faced in existing power marketing management, such as inaccurate load forecasting, delayed marketing decision-making, and low data quality.

[0024] As attached Figure 1 As shown, the specific situations of the three embodiments in data collection, preprocessing, feature extraction, load forecasting and marketing decision-making will be analyzed in detail below.

[0025] Example 1: Small commercial area power marketing management application 1. Data collection and preprocessing In a small commercial area, such as a commercial street with multiple shops and small office buildings, voltage sensors, current sensors and power factor sensors are deployed. These sensors are connected through the sensor interface of the data acquisition module and collect power data every 10 minutes. The data is collected continuously for two months (assuming 30 days per month), with a total of n=8640 times. The collected data includes the power consumption of each shop and office building in different time periods. For example, the power parameters during the peak business hours during the day and the rest period at night are obviously different.

[0026] Calculate the average value of voltage, current, and power factor V av e 、I av e 、PF av e and standard deviation s .set up k 1 =3, k 2 =3, k 3 =0.5, and the collected data is judged and removed for abnormal values. For example, a store has an abnormal voltage fluctuation due to a momentary electrical failure, and the abnormal value is successfully identified and removed. At the same time, for a small number of missing values ​​caused by a brief interruption in sensor transmission, the average value of adjacent data is used to fill in the missing values, and the preprocessed data D' is obtained.

[0027] 2. Feature extraction and load forecasting Calculate the power consumption characteristic value E based on the preprocessed data D', and initially set the collection time interval T = 10 minutes, adjust parameters α =0.5, β =0.1, dynamically adjust the acquisition frequency △t according to the real-time total power of the commercial area. For example, during holiday promotions, the total power of the commercial area increases, and the acquisition frequency is automatically accelerated to obtain more accurate power consumption characteristics. At the same time, the power load fluctuation characteristic value F is calculated.

[0028] When constructing the load forecasting model, the number of input layer neurons is determined to be 6 according to the dimensions of the power consumption characteristic value (including voltage, current, power factor, timestamp, acquisition frequency and store type identification dimensions), the number of output layer neurons is 1, and the adjustment coefficient k=2. The number of hidden layer neurons is calculated The historical power data of the commercial area in the past year was collected and processed according to the above data collection and preprocessing steps to train the neural network model. The weights and biases were adjusted using the back propagation algorithm. The mean square error (MSE) was used as the loss function. The training was performed until the loss function value converged to about 0.6, and a suitable load forecasting model was obtained.

[0029] 3. Marketing decision and implementation The power consumption characteristic value calculated in real time E Input the trained load forecasting model to get the predicted power load value L Assuming the preset high load threshold L=500kW, the electricity price adjustment coefficient c =0.1, when the predicted load value L =600kW (such as during the peak period of air conditioning use in summer), the increase in peak electricity prices T h =0.1×(600-500)=10 yuan / MWh, and increase the peak electricity price accordingly to encourage shops to use electricity at off-peak times. Preset low load threshold L 1 =100kW, electricity price adjustment coefficient d = 0.05, when the predicted load value L =80kW (such as during non-business hours at night), the reduction in electricity prices during off-peak hours =0.05×(100-80)=1 yuan / MWh, reducing off-peak electricity prices and attracting some commercial activities that can flexibly arrange electricity usage time to be carried out during off-peak hours.

[0030] In terms of power resource allocation, assuming that the commercial area is divided into three main areas (Area A is a catering area, Area B is a retail area, and Area C is an office area), the predicted load value of Area A isL 1 =200kW, predicted load value for area B L 2 =150kW, predicted load value for area C L 3 =100kW, resource allocation adjustment factor i =0.3, calculate the allocation ratio of each area according to the formula, give priority to the catering concentration area (Area A) and other areas with high load values ​​and obvious growth trends (such as during peak meal times) to ensure the normal operation of commercial activities. If the preset promotion load range is [300kW, 400kW], the discount adjustment coefficient e =0.08, when the predicted load value L =350kW (such as during joint promotions by merchants), calculate the discount d =0.08×(350-300)=4 yuan, push corresponding preferential packages to users, stimulate consumption, and increase electricity sales.

[0031] Example 2: Electricity marketing management application in large industrial plants 1. Data collection and preprocessing In a large industrial plant, covering multiple production workshops and large equipment, various types of sensors were installed to collect power data every 8 minutes for four consecutive months. The collected data reflects the power consumption under different production processes, different shifts, and equipment maintenance cycles. The relevant mean and standard deviation are calculated and set k 1 =2.5, k 2 =2.5, k 3 =0.4, outlier processing and missing value filling are performed on the data. For example, the current spike generated at the moment of startup of a large device is identified as normal fluctuation and is not misjudged as an outlier; and individual missing values ​​caused by sensor aging are accurately filled by the weighted average method of adjacent data (different weights are assigned according to the time interval between adjacent data and missing values), obtaining reliable preprocessed data. D' .

[0032] 2. Feature extraction and load forecasting Calculate the power consumption characteristic value E according to D', the initial collection time interval T = 8 minutes, adjust parameters α =0.6, β =0.12, dynamically adjust the acquisition frequency according to the real-time total power of the plant △tFor example, during the trial production of new products, equipment debugging causes large power fluctuations, and the acquisition frequency is automatically adjusted to obtain more detailed power consumption characteristics. At the same time, the power load fluctuation characteristic value F is calculated.

[0033] When building the load forecasting model, the number of input layer neurons is determined to be 7 (including voltage, current, power factor, timestamp, acquisition frequency, workshop type identification and equipment operation status dimensions), the number of output layer neurons is 1, the adjustment coefficient k=2.5, and the number of hidden layer neurons is calculated. The historical power data of the plant area in the past three years were collected and processed for model training. The back propagation algorithm was used, and the mean square error (MSE) was used as the loss function. The training was carried out until the loss function value converged to about 0.4, and an accurate load forecasting model was obtained.

[0034] 3. Marketing decision and implementation The calculated power consumption characteristic value E is input into the trained model to obtain the predicted power load value L . Assume that the preset high load threshold L 1 =10000kW, electricity price adjustment coefficient c =0.12, when the predicted load value L =11000kW (such as during overtime work to meet orders during the peak production season), the increase in peak electricity prices T h =0.12×(11000-10000)=120 yuan / MWh, adjust the electricity price to encourage enterprises to optimize production plans and reduce electricity load during peak hours. Preset low load threshold L 1 =3000kW, electricity price adjustment coefficient d =0.08, when the predicted load value L =2500kW (such as during equipment maintenance), the reduction in off-peak electricity prices =0.08×(3000-2500)=40 yuan / MWh, encouraging enterprises to arrange some non-emergency production tasks during off-peak hours.

[0035] In terms of power resource allocation, assuming that the factory area is divided into five main areas (Area A is the foundry workshop, B is the processing workshop, C is the assembly workshop, D is the storage area, and E is the office area), the predicted load value of area A is L 1 =4000kW, predicted load value for area B L 2 =3000kW, predicted load value for area C L 3 =2000kW, predicted load value for area D L 4=500kW, predicted load value for area E L 5 =300kW, resource allocation adjustment factor i =0.4, calculate the allocation ratio of each area according to the formula, give priority to the power supply of the foundry (Area A) and other areas with high load values ​​and obvious growth trends (such as the batch production stage of new products) to ensure smooth production. If the preset promotion load range is [6000kW, 8000kW], the discount adjustment coefficient e =0.1, when the predicted load value L =7000kW (such as when launching a new production process with significant energy-saving effect), calculate the preferential strength d =0.1×(7000-6000)=100 yuan, push customized electricity package discount information to enterprises, encourage enterprises to adopt energy-saving production technologies, and at the same time improve the power sales volume and market competitiveness of power companies.

[0036] Example 3: Electricity marketing management application in residential areas 1. Data collection and preprocessing In a residential area, sensors were installed at the distribution boxes of each building to collect power data every 15 minutes for six months. The collected data covered different power usage scenarios in residents' daily lives, such as power parameters during daytime working hours, peak power usage hours at night, and special periods such as holidays. The average values ​​of voltage, current, and power factor were calculated. V av e 、I av e 、PF av e and standard deviation s .set up k 1 =3.5, k 2 =3.5, k 3 =0.6, outlier processing and missing value filling are performed on the data. For example, voltage fluctuations caused by the instantaneous start-up of high-power electrical appliances in individual residents’ homes are correctly identified and do not affect data quality; for a small number of missing values ​​caused by network transmission problems, the average value of adjacent data is used for effective filling to obtain preprocessed data D' .

[0037] 2. Feature extraction and load forecasting Calculate the power consumption characteristic value E according to D', the initial collection time interval T = 15 minutes, adjust parameters α =0.4, β=0.08, dynamically adjust the acquisition frequency according to the real-time total power of the cell △t For example, during the peak period of air conditioning use in summer, the total power increases and the acquisition frequency is accelerated. At the same time, the power load fluctuation characteristic value F is calculated.

[0038] When constructing the load forecasting model, the number of neurons in the input layer is determined to be 5 according to the dimension of the power consumption characteristic value (including voltage, current, power factor, timestamp, and acquisition frequency), the number of neurons in the output layer is 1, and the adjustment coefficient k =3, calculate the number of neurons in the hidden layer The historical power data of the community in the past two years were collected and processed according to the above process to train the neural network model. The weights and biases were adjusted using the back propagation algorithm. The mean square error (MSE) was used as the loss function. The training was performed until the loss function value converged to about 0.5, and an effective load forecasting model was obtained.

[0039] 3. Marketing decision and implementation The power consumption characteristic value E calculated in real time is input into the trained load forecasting model to obtain the predicted power load value L. Assuming the preset high load threshold L =2000kW, electricity price adjustment coefficient c =0.08, when the predicted load value L =2200kW (e.g., during the peak evening electricity consumption period in summer), the increase in peak electricity prices T h =0.08×(2200-2000)=16 yuan / MWh, by appropriately raising electricity prices to guide residents to use electricity during off-peak hours, and relieve the pressure on the power grid. Preset low load threshold L 1 =500kW, electricity price adjustment coefficient d =0.03, when the predicted load value L =400kW (e.g. in the early morning), the reduction in off-peak electricity prices =0.03×(500-400)=3 yuan / MWh, encouraging residents to use electric water heaters and other electrical appliances that can be flexibly arranged during off-peak hours to improve the utilization rate of power equipment.

[0040] In terms of power resource allocation, assuming that the community is divided into three areas (area A is multi-story residential, area B is high-rise residential, and area C is the community public facilities area), the predicted load value of area A is L 1 =1000kW, predicted load value of area B L 2 =800kW, predicted load value for area C L 3=200kW, resource allocation adjustment factor θ=0.2, calculate the allocation ratio of each area according to the formula, give priority to multi-storey residential buildings (Area A) and other areas with high load values ​​and obvious growth trends (such as the evening when residents return home in large numbers), to ensure that residents' lives are not affected. If the preset promotion load range is [1500kW, 1800kW], the discount adjustment coefficient e = 0.05, when the predicted load value L =1600kW (e.g. during the period when power companies promote smart home electricity packages), calculate the discount d =0.05×(1600-1500)=5 yuan, and push preferential information of smart home electricity packages to residents, such as electricity fee reduction for purchasing smart meters, etc., to encourage residents to adopt smart electricity equipment, improve electricity efficiency, and increase user stickiness of power companies. In addition, the user feedback receiving unit in the marketing decision module will collect residents' feedback on strategies such as electricity price adjustment and promotion activities. For example, if residents reflect that the increase in peak electricity prices in summer is too large and affects the cost of living, power companies can adjust strategies appropriately based on feedback, optimize marketing decisions, and further improve residents' satisfaction.

[0041] Through the above embodiments of different scenarios (small commercial areas, large industrial plants, and residential areas), it can be seen that the present invention ensures high data quality in the data collection and preprocessing stages through reasonable sensor deployment, accurate outlier processing, and missing value filling. In the process of feature extraction and load forecasting, the acquisition frequency is dynamically adjusted according to the actual operating conditions of the power system, and the powerful nonlinear fitting ability of the neural network model is used to effectively improve the accuracy of load forecasting and the adaptability to load fluctuations. Finally, in the marketing decision-making link, the diversified strategies formulated according to the prediction results, whether it is electricity price adjustment, resource allocation, or promotional activities, have achieved the optimal allocation of power resources, improved the economic benefits and service quality of the enterprise, and fully demonstrated the effectiveness and practicality of the present invention in the power marketing management system.

[0042] As attached Figure 2 As shown, the following introduces a system architecture and module functions, system operating environment, system implementation form and other related contents for a power marketing management system, so that those skilled in the art can implement it.

[0043] (1) System architecture and module functions. The system is mainly composed of data acquisition module, data preprocessing module, feature extraction module, load forecasting module, marketing decision module, data storage module and communication module.

[0044] The data acquisition module collects data such as voltage, current, power factor and timestamp of each node of the power system in real time through a variety of sensor interfaces. The data preprocessing module processes outliers and fills missing values ​​in the collected data, calculates statistics and judges outliers based on thresholds. The feature extraction module calculates the characteristic values ​​of power consumption and load fluctuation based on the preprocessed data, and dynamically adjusts the acquisition frequency based on the system power. The load forecasting module adopts a neural network model, whose structure is determined based on the characteristic value of power consumption. It is trained through historical data and uses the mean square error as the loss function until it converges to a suitable range. The marketing decision-making module formulates electricity price adjustment, resource allocation and promotion activity strategies based on the predicted load value, and optimizes according to user feedback. The data storage module stores various types of data, and the communication module realizes the data and command transmission between internal and external devices in the system.

[0045] After the data acquisition module acquires the power data, it is transmitted to the data preprocessing module for processing to obtain reliable preprocessed data. The feature extraction module calculates the eigenvalue based on the preprocessed data, and the load forecasting module inputs the eigenvalue into the trained neural network model to predict the power load value. The marketing decision module formulates strategies based on the predicted values, the data storage module stores relevant data, and the communication module ensures the transmission of data and instructions. Each module works together to realize the full process of power marketing management.

[0046] (2) System operating environment Hardware environment: The system needs to be equipped with a processor of certain performance to perform tasks such as data processing, model calculation and decision making. At the same time, sufficient capacity of memory is required, including volatile memory (such as RAM for data storage and operation during program operation) and non-volatile memory (such as ROM for storing system programs and fixed data) to store system programs, data and temporary data during operation. In addition, various sensors (such as voltage sensors, current sensors and power factor sensors) need to be connected to the data acquisition module to realize the collection of power data.

[0047] Software environment: The operating system provides a basic platform for system operation, manages hardware resources and provides an operating environment for other software. The system contains one or more applications, such as data acquisition programs, data preprocessing programs, load forecasting programs, and marketing decision-making programs, to implement the various functions of the system. At the same time, the system also relies on software components such as database management system (used by data storage module to manage data storage and retrieval), communication protocol software (used by communication module to realize data communication) and deep learning framework (used by load forecasting module to build and train neural network models).

[0048] (3) System implementation form The power marketing management system of the present invention is implemented in the form of a combination of software and hardware. The hardware part includes physical devices such as sensors, processors, and memories, which provide the basis for data collection, calculation, and storage for the system. The software part includes operating systems, applications, and various drivers, etc., which realize the logical functions of the system, such as data processing, analysis, prediction, and decision-making. The software runs on the hardware platform and interacts with the hardware devices through the driver to realize the overall functions of the system.

[0049] The system exists in the form of a computer program product, and the computer program is stored on a computer-usable storage medium (such as disk storage, CD-ROM, optical storage, etc.). These programs contain processor executable instructions for implementing various functional modules of the system, such as data acquisition, preprocessing, feature extraction, load forecasting, and marketing decisions. By installing the computer program on the target device, the device is equipped with the functions of a power marketing management system, realizing effective management and decision support for power marketing activities. At the same time, the system can also be customized and expanded according to actual needs to meet the business needs of different power companies.

[0050] In order to verify the impact of data preprocessing on the performance of the load forecasting model, we also conducted simulation experiments on the MFNN, LSTM, and CNN models.

[0051] 1. Experimental Preparation 1. Dataset selection The power load data spanning one year is collected from a regional power system and recorded at hourly intervals, with a total of about 8760 sample points (365×24). In addition to the power load value, auxiliary parameter data such as voltage, current, and power factor are collected synchronously to construct a multi-dimensional data set to fully reflect the power consumption characteristics and system operation status. This data set can simulate the actual power supply and demand scenarios, provide rich information for model training, enable the model to learn the law of power load changes and the intrinsic relationship with related parameters, ensure the model's prediction performance reliability and generalization ability under complex power conditions, provide accurate data support for power marketing management decisions, and improve the efficiency of power resource allocation and the stability of power grid operation.

[0052] 2. Experimental environment construction Based on the Python language ecosystem, the PyTorch deep learning framework is integrated to build an experimental environment. The multi-layer feedforward neural network MFNN is used as the basic architecture, and the input layer neurons are configured according to the characteristic dimensions of the power load data. After data exploratory analysis, it is determined to be m neurons, accurately receiving multiple parameter inputs such as voltage, current, and power factor. There are two hidden layers, and the number of neurons is determined by empirical formulas and repeated experiments. h 1 andh 2 ,use ReLU The activation function introduces nonlinear factors to enhance the model's ability to learn complex power load characteristics. The output layer single neuron output predicts the power load value, and the mean square error is used as the loss function to measure the prediction error. The stochastic gradient descent algorithm optimizes the model parameters. The learning rate is initially set to 0.01 and dynamically adjusted according to the training convergence situation to ensure efficient and stable model training, accurately capture the power load fluctuation pattern, provide accurate prediction tools for power marketing management, optimize power generation plans, reduce operating costs, improve power service quality, and enhance the market competitiveness and economic benefits of power companies.

[0053] The data set is divided into training set, validation set and test set in a ratio of 7:2:1, with about 6132 samples in the training set, 1752 samples in the validation set and 876 samples in the test set. Stratified sampling ensures that each data set retains the original distribution characteristics of the data, such as the load change patterns in different seasons, weekdays and holidays, and improves the generalization ability of the model under various working conditions. The training set is used for the model to learn the relationship between load change patterns and parameter mapping; the validation set monitors the training process, evaluates the rationality of the model architecture and parameter settings, and prevents overfitting; the test set simulates real application scenarios, fairly evaluates the model's prediction performance on unknown data, ensures the reliability of the model's transition from training to actual application, and establishes a rigorous and scientific process for the performance evaluation of the power load forecasting model, ensures the stable and efficient operation of the model in the power marketing management system, improves the scientificity and accuracy of power resource allocation, and promotes the upgrade of intelligent operation and management in the power industry.

[0054] 3. Experimental Grouping The experiment set up two groups for comparison: the group before data preprocessing directly used the original data to train MFNN, LSTM, and CNN models; the group after data preprocessing trained the same model after processing the data according to the invented method. The training and evaluation of each group of models are independent to ensure the comparability and fairness of the experimental results. Through comparative analysis, the performance improvement effect of data preprocessing on different models and the differences in applicable scenarios are clarified, providing a key basis for the power marketing management system to select the optimal model and data processing strategy, improve the accuracy and efficiency of power load forecasting, optimize power production and operation scheduling, realize the economic and reliable operation of the power system, meet the society's growing power demand, and promote the implementation of the sustainable development strategy of the energy industry. 2. Experimental steps 1. Model training and evaluation before data preprocessing (Group A) Data processing: The original data set skips the abnormal value processing and missing value filling process and directly normalizes the data. For parameters such as voltage, current, and power factor, the formula Map the values ​​to the interval [0,1], x norm is the normalized value, x is the original value, x min and x max The normalization balances the difference in data dimensions, accelerates the convergence of model training, improves training stability and accuracy, ensures that the model pays equal attention to learning of each parameter, avoids learning deviation due to different dimensions, and lays the foundation for the model to accurately capture the complex nonlinear relationship between power load and parameters, and improves the upper limit of the model's load prediction performance under unprocessed data. Although it cannot cure the problem of data quality, it provides a benchmark reference for comparing preprocessing effects, helps to quantitatively evaluate the value contribution of the preprocessing link, and optimizes the power load prediction process and model architecture design.

[0055] Model training: The normalized training data of Group A is input into the MFNN model. The number of training rounds is set to 100. The SGD optimization algorithm iteratively updates the model weights. The learning rate is 0.01 to control the update step size. The validation set loss value is calculated every 10 rounds ( loss val1 ) and accuracy( acc val1 ) monitor the training progress. The loss value reflects the degree to which the prediction deviates from the true value, and the accuracy measures the proportion of samples predicted correctly. In the early stage, the loss value is high and the accuracy is low. As the training progresses, the model learns the load characteristics, the loss value decreases, and the accuracy increases. If the loss value of the validation set increases instead of decreasing, or the accuracy stagnates, it indicates that the model is overfitting, and it is necessary to adjust the parameters or improve the architecture to ensure the balance between the generalization ability of the model and the prediction accuracy, provide a reliable load forecasting model for power marketing management, optimize the formulation of power generation plans, improve the economic and safe operation of the power grid, and adapt to the dynamic changes in the power market.

[0056] Model evaluation: After training, the MFNN model is evaluated using the Group A test set and the loss value is calculated ( loss test1 ) and accuracy( acc test1 ). The loss value measures the average level of the model's prediction error for unknown data, and the accuracy reflects the proportion of samples accurately predicted by the model in the test set, intuitively showing the actual prediction performance of the model. The test set is independent of the training validation set, and the results truly reflect the generalization and prediction accuracy of the model, providing key quantitative indicators for evaluating the feasibility of the model in the power marketing management system. Comparing the test performance of different models helps to select the optimal model, improve the accuracy and timeliness of power load forecasting, optimize power resource allocation management, reduce operating costs, enhance the market competitiveness and service quality of power companies, and promote the intelligent digital transformation of the power industry.

[0057] 2. Model training and evaluation after data preprocessing (Group B) Data processing: Group B processes data according to the invented method, first counting the mean and standard deviation of voltage, current, and power factor, and setting the threshold ( k 1 ,k 2 ,k 3 ) to determine abnormal values. Taking voltage as an example, if (V is the voltage value, and The missing values ​​are filled with the mean of adjacent data to maintain the continuity and integrity of the data, and then normalized. This preprocessing accurately eliminates abnormal interference, makes up for missing information, improves data quality and reliability, creates high-quality data conditions for the model to learn the essential laws of power load changes, improves the quality of model input from the root, significantly enhances the stability and accuracy of model prediction performance, accurately captures the dynamic change trend of load, provides a solid data foundation for power marketing decisions, supports the refined management and operation of power companies, and improves the overall operation efficiency, benefits and reliability of the power system.

[0058] Model training: The MFNN model is trained on the processed data of Group B, and the architecture and training parameters are the same as those of Group A. The validation set loss value is recorded every 10 rounds during training ( loss val2 ) and accuracy( acc val2 ). Due to the excellent data quality, the model learning is efficient and stable, the loss value decreases rapidly, and the accuracy rate increases steadily, reflecting the positive driving effect of preprocessing on model training, so that the model can accurately fit the relationship between power load and parameters, explore potential change patterns, and provide a high-precision load forecasting model for power marketing management, helping enterprises to optimize power dispatching strategies, improve power supply reliability, reduce operating costs and energy consumption emissions, achieve economic, efficient, green and sustainable development of the power system, and enhance the market competitiveness and social responsibility of power companies.

[0059] Model evaluation: Use the test set after Group B processing to evaluate the MFNN model and calculate ( loss test2 ) and accuracy( acc test2 ). With the support of high-quality data, the model test performance is excellent, with low loss value and high accuracy. Compared with Group A, the preprocessing quantitatively reflects the effect of improving the prediction performance of the MFNN model, verifies the key value of the preprocessing link, and provides a strong practical basis for its promotion and application in power marketing management systems.

[0060] 3. LSTM and CNN model experimental process The LSTM and CNN models repeat the above MFNN model experimental process in Group A and Group B. LSTM processes the long-term and short-term dependencies of time series power load data with its unique gating mechanism, and CNN uses convolutional layers to capture local features and global patterns of data. The two have different performances in different data quality scenarios. By comparing the experimental results of each model in the two groups, we deeply analyze the differences in the impact mechanism of data preprocessing on different model architectures, explore the advantages and applicable scenarios of each model, and provide scientific decision-making basis for the construction of hybrid models or the selection of the optimal single model for the power marketing management system, so as to achieve the coordinated improvement of power load forecasting accuracy, efficiency and stability, promote the intelligent upgrading and innovative development of power marketing business, and meet the diverse and complex needs of the power market and the requirements of the industry's sustainable development strategy.

[0061] 3. Experimental Results Analysis 1. Before data preprocessing: In Group A, the validation set loss value of the MFNN model is 0.85 (mean square error), which means that there is a large deviation between the prediction results of the model on the validation set and the true value. The main reason is that the outliers and missing values ​​in the original data interfere with the model's learning process of the complex relationship of power load, making it difficult for the model to accurately fit the data distribution law, resulting in violent fluctuations in the prediction results and high errors. The test set loss value reached 0.92, further indicating that the generalization ability of the model on unknown data is severely limited and cannot be effectively promoted and applied to actual power marketing management decision-making scenarios. The validation set accuracy is 78.00%, and the test set accuracy is 73.00%. A large number of sample prediction errors have a great negative impact on the reliability of power load forecasting, which can easily lead to significant deviations between power generation plans and power grid dispatching, thereby increasing operating costs and reducing power supply quality. This situation highlights the urgency and necessity of data preprocessing for improving model performance, and clarifies the key optimization directions and breakthrough points for the subsequent precise and intelligent development of power load forecasting models.

[0062] Compared with the MFNN model, the LSTM model has a validation set loss value of 0.82, a test set loss value of 0.90, a validation set accuracy of 79.00%, and a test set accuracy of 74.00%. Although LSTM can handle long-term and short-term dependencies of time series power load data with its unique gating mechanism, the quality of the original data still has a significant impact on its performance. It is insufficient in capturing data features and making accurate predictions, and cannot give full play to its architectural advantages.

[0063] The CNN model has a validation set loss of 0.83, a test set loss of 0.91, a validation set accuracy of 78.50%, and a test set accuracy of 73.50%. Although CNN can use convolutional layers to capture local features and global patterns of data, it is difficult to accurately extract effective features under the interference of raw data, resulting in poor prediction accuracy and failing to meet expectations in actual power load forecasting tasks.

[0064] 2. After data preprocessing: After data preprocessing, the performance of the MFNN model of Group B has significantly improved. The loss value of the validation set has dropped sharply to 0.45, and the loss value of the test set has dropped to 0.55. This is due to the fact that the preprocessing operation accurately removes outliers and effectively fills missing values, allowing the model to accurately capture the core characteristics and changing trends of the power load, greatly reducing the prediction error. The accuracy of the validation set climbed to 86.00%, and the accuracy of the test set increased to 83.00%. The proportion of correct samples predicted by the model in the validation and actual tests has increased significantly, and the credibility of the prediction results has been greatly enhanced, which can provide solid and reliable support for accurate decision-making in power marketing management. For example, when formulating power generation plans, power demand can be more accurately estimated to avoid over-generation or under-generation; in the power grid dispatching link, power resources can be reasonably allocated based on accurate prediction results, and the economic operation level of the power grid can be comprehensively improved, which effectively demonstrates the core value and key effectiveness of data preprocessing in improving the performance of the MFNN model and promoting the transformation of intelligent and efficient operation and management in the power industry, and builds a strong driving force for the digital transformation and sustainable development of power companies.

[0065] The validation set loss value of the LSTM model dropped to 0.50, the test set loss value was 0.62, the validation set accuracy was 84.00%, and the test set accuracy was 81.00%. Although data preprocessing improves the performance of the LSTM model, it is not as good as the MFNN model in terms of loss reduction and accuracy improvement. After analysis, under the experimental data characteristics and load forecasting task scenario, the MFNN model structure can more efficiently learn the intrinsic relationship between power load and parameters after preprocessing and optimizing data input. Its parameter adjustment strategy performs better when processing such data and contributes more to improving prediction accuracy and stability. Therefore, the MFNN model is more suitable in the load forecasting model of the present invention.

[0066] The validation set loss value of the CNN model dropped to 0.48, the test set loss value was 0.60, the validation set accuracy was 84.50%, and the test set accuracy was 82.00%. Although the performance of the CNN model has also been significantly improved, when processing the power load data, the MFNN model has more advantages in capturing the overall trend and the comprehensive balance of detailed features. After preprocessing, the MFNN model can better exert the multi-layer connection learning ability of neurons with the optimized data, and achieve better prediction performance with a simpler architecture, which is more in line with the load prediction demand scenario of the present invention.

[0067] 4. Experimental Conclusion In summary, from the analysis of experimental results combined with the Figure 3 , we draw the following conclusions: (1) Data preprocessing plays a key role in improving the performance of MFNN, LSTM, and CNN models, significantly reducing the model loss value, greatly improving the accuracy, effectively overcoming the defects of the original data, laying a solid foundation for the accurate prediction of the power marketing management system, optimizing the power resource allocation management process in an all-round way, comprehensively improving the power industry's operational efficiency, economic benefits, market competitiveness and sustainable development capabilities, and strongly promoting the power industry to move towards a new journey of intelligent and refined management. (2) After experimental comparison, the MFNN model performs excellently in the load forecasting task of the present invention, and is more suitable for processing preprocessed power load data. It has significant results in reducing loss values ​​and improving accuracy. After preprocessing and optimizing data input, it can deeply mine the potential characteristics and laws of the data with its own architectural advantages, accurately construct the mapping relationship between power load and multiple parameters, and stably output high-precision prediction results under different power conditions, providing solid technical guarantees for accurate decision-making and efficient operation of power companies, and leading the new direction of power load forecasting technology innovation and practical application development. (3) Comprehensively and in-depth analyze the performance characteristics of the MFNN model under different power conditions and massive data scales; continue to improve the power load forecasting technology system and practical methodology system, comprehensively and continuously improve the intelligent decision-making level of the power marketing management system, and inject lasting innovative vitality and solid technical support into the high-quality development of the power industry. By improving data quality and deeply mining features, the model learning efficiency and generalization ability are significantly enhanced, ensuring that the model operates stably and reliably in the actual application of power marketing management.

[0068] In order to fully verify that each module can be used in the power marketing management system, ensure the accuracy, reliability, effectiveness and stability of the system in actual application, and provide strong support for the marketing management decision-making of power companies. Figure 4 to 9 As shown, we tested each module using the above test cases, and all the results were qualified.

[0069] It can be seen that the power marketing management system designed by the present invention has excellent data processing capabilities, accurate and reliable data collection, efficient and effective preprocessing, and precise feature extraction with strong adaptability; the load forecasting accuracy is high and the scenario adaptability is wide, short-term forecasting helps daily operations, and long-term forecasting grasps trends; marketing decisions are scientific and reasonable, electricity price adjustments guide user behavior, resource allocation optimization protects key areas, and promotional activities stimulate electricity demand; the overall performance of the system is stable and reliable, the functions are well coordinated to achieve full process automation, it operates stably in complex environments with fault tolerance, and the performance indicators meet business needs to ensure efficient service, which is of great significance to power companies in improving marketing management and economic benefits.

[0070] It should be noted that through simulation experiments on the performance of multilayer feedforward neural network (MFNN), long short-term memory network (LSTM), and convolutional neural network (CNN) in load forecasting of power marketing management systems, we determined that the most suitable model for the present invention is preferably a multilayer feedforward neural network (MFNN). Of course, those skilled in the art can also use convolutional neural network (CNN) or long short-term memory network (LSTM) as equivalent replacements, which should be regarded as the scope of protection of the present invention.

[0071] 1. Experimental Preparation 1.1 Dataset Construction The power data of a regional power system in the past two years was collected, totaling 17,520 records. The region contains a variety of power consumption scenarios, such as industrial production, commercial operations, and residential electricity consumption, and changes in different seasons, working days, and holidays, making the data highly representative. After data preprocessing, removing obvious outliers and a small number of missing values, the final sample set was formed and divided into a training set (12,264 records), a validation set (3,504 records), and a test set (1,752 records) in a ratio of 7:2:1.

[0072] (II) Experimental environment construction Based on Python 3.8 programming language, combined with TensorFlow 2.5 and Keras 2.4.3 deep learning frameworks, the model is built and trained. The experimental hardware uses Intel Core i7-10700K processor, 32GB memory and NVIDIA GeForce RTX 3060 graphics card to accelerate the calculation process and ensure efficient execution of model training and evaluation.

[0073] (III) Model parameter setting MFNN: Taking a small commercial area as an example, the input layer sets 6 neurons according to the dimension of power consumption characteristic value (including voltage, current, power factor, timestamp, acquisition frequency and store type identification dimension); the output layer has 1 neuron outputting the predicted load value; the hidden layer has 2 layers, according to the formula " ”( m =6, i =1, k =2) The number of neurons in the first layer is calculated to be 5 and the number of neurons in the second layer is 4. The activation function is ReLU , optimizer selection Adam , learning rate is 0.001, loss function is mean square error (MSE), and batch size is 64.

[0074] LSTM: The input layer has 30 neurons according to the data feature dimension and time step; the hidden layer has 2 layers, each with 64 neurons, and uses the tanh activation function; the output layer has 1 neuron to output the predicted load value, and the optimizer is Adam , learning rate 0.001, loss function MSE , batch size 64, and number of training rounds 100.

[0075] CNN: The input layer has 30 neurons according to the data feature dimension and time step; the convolution layer has 2 layers, the first layer has 3×3 convolution kernels, 32 in number, and 1 in step; the second layer has 3×3 convolution kernels, 64 in number, and 1 in step; the activation function ReLU ; There are 2 pooling layers, both of which are maximum pooling, and the pooling window is 2×2; there are 2 fully connected layers, with 128 neurons in the first layer and 64 neurons in the second layer, and the last output layer has 1 neuron to predict the load, the optimizer Adam , learning rate 0.001, loss function MSE, batch size 64.

[0076] 2. Experimental steps 1. Data preprocessing The original data is normalized so that each eigenvalue is in the interval [0,1], and the MinMaxScaler function is used to achieve it. The formula is: ,in x is the original value, x min and x max are the minimum and maximum values ​​of the feature.

[0077] 2. Model training The training set is used to train MFNN, LSTM, and CNN models respectively. The number of training rounds is set to 100. The training loss and validation loss are recorded in each round. Early stopping is performed according to the validation loss (training is stopped if the validation loss does not decrease for 10 consecutive rounds) to prevent overfitting and obtain the optimal model parameters.

[0078] 3. Model evaluation Evaluate the trained model using the test set and calculate the mean squared error ( MSE ), mean absolute error ( MAE ) and the mean absolute percentage error ( MAPE ) measures the prediction accuracy, and the formula is as follows:

[0079] Where n is the number of test samples, is the actual load value, is the predicted load value.

[0080] 3. Experimental Results Analysis As attached Figure 10 to 11 As shown, MFNN MSE、MAE and MAPE The value is relatively the lowest among the three, indicating that the deviation between its predicted value and the true value is the smallest and the prediction accuracy is the highest.

[0081] The training time of MFNN is 2.5 hours, LSTM is 4.2 hours, and CNN is 3.8 hours. MFNN has a relatively simple structure, few training parameters, and the shortest training time. It meets the real-time requirements of the power marketing management system and can quickly adjust the prediction model according to new data to ensure the timeliness of load prediction and provide timely support for marketing decisions.

[0082] MFNN has an intuitive structure, and the relationship between the number of layers and neuron connections is easy to understand, which is convenient for technicians to adjust and optimize. It can flexibly modify the network structure and parameters based on power marketing expertise and actual needs, so that the model and business are deeply integrated. However, LSTM and CNN have complex internal structures, many parameters, weak interpretability, and difficult adjustments, which increase the cost and difficulty of application and maintenance.

[0083] 4. Experimental Conclusion In terms of comprehensive prediction accuracy, training time, model complexity and interpretability, the multi-layer feedforward neural network performs excellently in the load forecasting task of the power marketing management system of the present invention. Its high-precision prediction capability provides a reliable basis for electricity price adjustment, resource allocation, and promotion strategy formulation; the short training time meets the real-time requirements of the system; the simple structure is convenient for technicians to operate and maintain, realize intelligent and efficient power marketing management, and improve the operating efficiency and market competitiveness of power companies. Therefore, MFNN is the best choice for the load forecasting model of the present invention.

[0084] In summary, compared with the existing traditional power marketing methods, the present invention has the following significant advantages: (I) Real-time data collection and transmission: In the data collection module, the sensor network and the system are seamlessly connected through the "Internet +". After each node sensor (voltage, current, power factor sensor) collects data, it is uploaded in real time through the Internet communication protocol. In small commercial areas, 8,640 data collected every 10 minutes, several months of data collected every 8 minutes in large industrial plants, and half a year of data collected every 15 minutes in residential areas can be accurately and instantly transmitted to the data center. This real-time capability enables power companies to control the operating status of the power system in real time, gain timely insight into power load fluctuations, and respond quickly during commercial power peaks, critical periods of industrial production, or alternating peaks and valleys in residential power consumption, thereby gaining an advantage in subsequent marketing decisions and avoiding decision-making errors and power supply risks caused by data lags.

[0085] (II) Implementation of precision marketing decisions: The electricity price adjustment strategy, resource allocation plan and promotion activity plan formulated based on the load forecast value are accurately pushed to the target users through the Internet. For example, the peak electricity price increase in small commercial areas, the off-peak electricity price discount in industrial plants and the promotion of smart electricity packages in residential areas are all reached to the user end in real time through the Internet. This precise push accurately locates user needs, improves the conversion rate of marketing activities, optimizes the allocation of power resources, enhances users' acceptance of electricity price adjustments, and improves corporate economic benefits and user satisfaction.

[0086] (III) Powerful computing power ensures data processing efficiency: In data preprocessing, when dealing with outlier screening and missing value filling tasks for massive power data, cloud computing distributed architecture performs parallel computing, which greatly reduces processing time. In the feature extraction stage, dynamic adjustment of acquisition frequency and complex feature value calculation are completed instantly with the support of cloud computing. Taking MFNN model training as an example, in the face of large-scale historical power data, cloud computing intelligently allocates computing resources according to demand, allowing the model to complete training in as short as 2.5 hours (compared to 4.2 hours for LSTM and 3.8 hours for CNN), ensuring that the model is quickly updated when the power load changes dynamically, and accurate predictions are always made, which strongly supports the timeliness of marketing decisions and enhances the market resilience of power companies.

[0087] (IV) Elastic resource allocation optimizes cost-effectiveness: During peak electricity consumption periods (such as high temperatures in summer when residential electricity consumption peaks or peak industrial production seasons), the system's computing capacity increases, and cloud computing automatically expands the scale of resources to ensure smooth data processing; during off-peak periods, resources are reduced to cut operating costs. This elastic resource management mechanism significantly optimizes the investment and operation and maintenance costs of enterprise IT infrastructure, improves resource utilization, and lays a solid foundation for the long-term stable operation of the enterprise, while ensuring the efficiency and stability of the power marketing management system.

[0088] (V) Comprehensive market insight and personalized strategy formulation: Through in-depth analysis of massive power data (covering power consumption data in multiple regions, multiple time periods, and multiple user types), we can gain comprehensive insight into market demand and user behavior patterns. In the commercial field, we can accurately grasp the peak and valley power consumption and power demand characteristics of shops in different formats; at the industrial level, we can plan the allocation of power resources based on workshop production processes and equipment operation cycles; at the residential end, we can customize electricity price packages and promotional activities based on daily routines and seasonal power consumption habits. This precision marketing strategy fully meets the personalized needs of users, optimizes the power consumption structure, improves the efficiency of power resource market allocation, and stimulates market vitality.

[0089] (VI) In the implementation of marketing decisions, big data continuously provides feedback to evaluate the effectiveness of decisions. By collecting user feedback on electricity price adjustments, resource allocation responses, and promotion participation data, the system conducts in-depth mining and analysis to accurately locate decision deviations and optimization space. For example, according to the feedback of users in residential areas on peak electricity prices, the adjustment range is appropriately fine-tuned, and resource allocation factors are optimized according to changes in the electricity consumption structure of industrial enterprises. Continuous iterative optimization enables power marketing decisions to closely follow market dynamics and the evolution of user needs, improves decision accuracy and foresight, strengthens corporate market competitiveness and industry leadership, and promotes the advancement of power marketing management towards intelligence and refinement.

Claims

1. An information processing method for a power marketing management system, characterized in that: The following steps are involved: S1 Data Collection: Collect power data through sensors installed at various nodes of the power system and mark the collected data as D ={(V1,I1,PF1,t1),( V 2 ,I 2 ,PF 2 ,t 2 ),…,( V n ,I n ,PF n ,t n )},in n is the number of collections, V is the voltage value, I is the current value, PF is the power factor, t is the timestamp; S2 data preprocessing: data collected by S1 D Preprocessing is performed, including removing outliers and missing values. The outlier judgment basis is that if the voltage value V satisfies |VV av e |>k 1 σ , or the current value I satisfies |II av e |>k 2 σ , or the power factor PF satisfies |PF- PF av e |>k 3 σ , it is determined as an outlier and removed, where V av e 、I av e PF av e are the average values ​​of voltage, current and power factor respectively. σ is the standard deviation, k 1 、k 2 、k 3 is the preset threshold; missing values ​​are filled with the average value of adjacent data to obtain the preprocessed data D' ; S3 feature extraction: based on the data preprocessed by S2 D' Calculate the power consumption characteristic value E , the calculation method is: ， in V i For the i The voltage value is collected. I i For the i The current value is collected. PF i is the power factor collected for the i-th time, t i For the i The timestamp of the acquisition, Δt The collection is performed once at a predetermined interval, i.e., the collection frequency; S4 load forecasting: The power consumption characteristic value E Input into the pre-trained load forecasting model, the predicted power load value L ; S5 Marketing decision: Formulate power marketing strategies based on the predicted power load value L, including adjusting electricity price strategies, power resource allocation strategies, and power promotion activity strategies.

2. The information processing method for a power marketing management system according to claim 1, characterized in that: In the S2 data preprocessing step, the standard deviation σ The calculation method is: ， in x i For each data value of voltage, current, and power factor in the data set, is the average of all data values ​​in the data set.

3. The information processing method for a power marketing management system according to claim 1, characterized in that: In the feature extraction step S3, the acquisition frequency is dynamically adjusted according to the load fluctuation characteristics of the power system, and the adjustment method of Δt is: ， in T is the initial set collection time interval, α、β To adjust the parameters, P is the total power of the power system currently monitored, P 0 is the preset power threshold.

4. The information processing method for a power marketing management system according to claim 1, characterized in that: In the S3 feature extraction step, it also includes calculating the power load fluctuation characteristic value F , the calculation method is: ， in n Represents the total number of times data is collected. i Indicates the sequential index of data collection, E i It is the power consumption characteristic value calculated from the i-th collected data.

5. The information processing method for a power marketing management system according to claim 1, characterized in that: In the S4 load forecasting step, the load forecasting model adopts a neural network model, which is any one of MFNN, LSTM, and CNN models.

6. The information processing method for a power marketing management system according to claim 5, characterized in that: The load forecasting model adopts the MFNN model, the number of neurons in the input layer is determined according to the dimension of the power consumption characteristic value, and the hidden layer includes at least one layer. h Determined according to the following formula: ， in m is the number of neurons in the input layer, l is the number of neurons in the output layer, k The output layer outputs the predicted power load value for the future time period. L .

7. The information processing method for a power marketing management system according to claim 6, characterized in that: The training method of the MFNN model comprises the following steps: S4.1 collect historical power data, process it according to the data collection and preprocessing steps described in claim 1, and obtain historical power consumption characteristic values ​​and corresponding historical actual power load values; S4.2 takes the historical power consumption characteristic value as input and the historical actual power load value as output, and uses the back propagation algorithm to adjust the weights and biases of the neural network model for training to minimize the loss function value. The mean square error MSE is used as the loss function during the training process. The loss function calculation formula is: ， in q is the number of training samples, For the j The actual power load value of the samples, For the j The predicted power load value of samples; S4.3 continuously repeats the training process of S4.2 until the loss function value converges to between 0.1 and 1, which means that the neural network model meets the required prediction ability.

8. The information processing method for a power marketing management system according to claim 1, characterized in that: In the S5 marketing decision step, the electricity price adjustment strategy includes: When the predicted power load value is greater than the preset high load threshold, the peak period electricity price is increased according to the following formula: ， in To increase the amplitude, γ is the electricity price adjustment coefficient, L To predict the power load value, To preset a high load threshold; When the predicted power load value is less than the preset low load threshold, the off-peak period electricity price is reduced according to the following formula: ， in To reduce the amplitude, δ is the electricity price adjustment coefficient, L To predict the power load value, The preset low load threshold.

9. The information processing method for a power marketing management system according to claim 1, characterized in that: In the S5 marketing decision step, the power resource allocation strategy includes: According to the predicted power load values ​​of different regions, power resources are preferentially allocated to areas with higher load values ​​and obvious growth trends, calculated according to the following formula: ， in p For the distribution ratio, L To predict the power load value, L i For the i The predicted power load value of each region, m is the number of regions, θ Assign adjustment factors to resources.

10. The information processing method for a power marketing management system according to claim 1, characterized in that: In the S5 marketing decision step, the electricity promotion activity strategy includes: When the predicted power load value is within the preset promotion load range, the power package discount information is pushed to the user, calculated according to the following formula: ， in d For preferential treatment, ε is the preferential intensity adjustment coefficient, L To predict the power load value, [L p1 ,L p2 ] is the preset promotion load range.

11. A system for power marketing management, characterized in that: include: Data acquisition module: comprising a plurality of sensor interfaces, which is compatible with different types of sensors including voltage sensors, current sensors and power factor sensors, and is used to execute the S1 data acquisition step in the information processing method according to claim 1; Data preprocessing module: used to execute the S2 data preprocessing step in the information processing method according to claim 1; Feature extraction module: used to perform the S3 feature extraction step in the information processing method according to claim 1; Load prediction module: comprising the neural network model as claimed in claim 5, used to perform the load prediction step S4 in the information processing method as claimed in claim 1; Marketing decision module: including an electricity price adjustment unit, an electricity resource allocation unit and an electricity promotion activity unit, used to execute the S5 marketing decision step in the information processing method described in claim 1.

12. The system for power marketing management according to claim 11, characterized in that: Also includes: Data storage module: used to store the collected raw power data, pre-processed data, calculated characteristic values ​​and prediction results; Communication module: used to realize data communication between the system and other equipment in the power system, including data transmission between the data acquisition module and the sensor, and command transmission between the marketing decision module and the power marketing execution equipment.

13. The system for power marketing management according to claim 11, characterized in that: The marketing decision module also includes a user feedback receiving unit, which is used to receive user feedback information on the power marketing strategy and adjust and optimize the marketing strategy according to the feedback information.

Images