A user load forecasting system based on adaptive learning of historical load data

Abstract

The present invention relates to the technical field of load prediction of power systems, and discloses a user load prediction system based on adaptive learning of historical load data, comprising: a data acquisition and preprocessing module, a feature extraction module, an adaptive learning model module, a prediction module, an error evaluation and feedback module, and a historical database. The user load prediction system based on adaptive learning of historical load data takes data from smart meters and power monitoring equipment into a historical database through the data acquisition and preprocessing module, selects appropriate deep learning and mechanical learning in the adaptive learning model module, can better capture the law of load changes through adaptive learning, automatically adjusts the learning rate according to the changes in historical load data, uses training data to train the model, and continuously optimizes the parameters of the model, which can achieve accurate prediction of user loads, thereby improving the accuracy of system prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of power system load forecasting, and in particular to a user load forecasting system based on adaptive learning of historical load data. Background Art

[0002] With the development of the power market and the construction of smart grids, through the research and extension of power monitoring in different scenarios of power generation, power supply and power consumption, as well as load operation and power balance scenarios, we have carried out preliminary demand collection, sorting, non-core code writing, function verification and other work based on modules such as special monitoring business of key user power, including completing five functions such as user insight, power-on monitoring, district and city analysis, and green evaluation of prosperity index level, to optimize production and operation, improve customer service quality and promote technological development. Accurate user load forecasting is of great significance to the planning, operation and dispatching of power systems. How to effectively mine and use existing information for adaptive learning to improve forecasting accuracy is still a challenge currently faced, such as:

[0003] Traditional forecasting systems do not make full and flexible use of historical load data. They often only focus on the time series characteristics of load data, but ignore other potential information contained in the data, such as the periodic changes in load, the differences in load characteristics of different user types, etc. Traditional load forecasting methods often rely on fixed models and parameters, which are difficult to adapt to the complexity and diversity of load changes. Historical load data contains rich information, and existing systems find it difficult to adaptively learn the historical load data of different users for further prediction.

[0004] In view of the above problems, it is urgent to carry out innovative design based on the original power system load forecasting system. Summary of the invention

[0005] The purpose of the present invention is to provide a user load forecasting system based on adaptive learning of historical load data, so as to solve the problem proposed in the above background technology that the traditional forecasting system does not make full and flexible use of historical load data, often only focuses on the time series characteristics of the load data, and ignores other potential information contained in the data, and the traditional load forecasting method often relies on fixed models and parameters, and the existing system is difficult to adaptively learn the historical load data of different users for further prediction.

[0006] To achieve the above object, the present invention provides the following technical solutions: a user load forecasting system based on adaptive learning of historical load data, comprising a data acquisition and preprocessing module, a feature extraction module, an adaptive learning model module, a forecasting module, an error evaluation and feedback module and a historical database;

[0007] The data acquisition and preprocessing module is composed of a data acquisition submodule that collects data through communication with data sources such as smart meters and power monitoring equipment connected to users, a data cleaning submodule that performs quality inspection and processing on the collected data, and a data standardization submodule that performs standardization processing on the cleaned data and normalizes data such as load values ​​to a specific interval;

[0008] The feature extraction module is composed of a time feature extraction submodule for extracting time-related features of load data to construct a time series feature vector, a meteorological feature association submodule for collecting meteorological data corresponding to load data and analyzing the correlation between meteorological factors and load, and a holiday feature identification submodule for extracting and identifying holiday information in historical data;

[0009] The adaptive learning model module is composed of a model selection and construction submodule for selecting a suitable model architecture according to the characteristics of user load data and forecasting requirements, a parameter optimization submodule for optimizing and adjusting the parameters of the model using an adaptive optimization algorithm, and a model update and adaptive adjustment submodule for real-time update and adaptive adjustment of performance indicators of the monitoring model during the forecasting process;

[0010] The prediction module is composed of a short-term prediction submodule for obtaining recent historical composite data for short-term prediction of user load, a medium-term prediction submodule for obtaining one month's historical load data for medium-term load prediction of the power system, and a long-term prediction submodule for long-term load prediction based on long-term trend analysis of historical data;

[0011] The error evaluation and feedback module is composed of an error calculation submodule for calculating the error between the predicted composite value and the actual load value of the prediction module, an error calculation submodule for performing feedback analysis on the error calculation result of the error calculation submodule and the prediction result, and a model adjustment and optimization submodule for adjusting parameters, structure and favor selection model according to the results of the feedback analysis;

[0012] The historical database is the main part of the user's electricity consumption historical load data record.

[0013] Historical data from a historical database.

[0014] By adopting the above technical solution, accurate prediction of user load can be achieved by deeply mining the complex patterns and potential laws in historical load data.

[0015] Preferably, the data acquisition submodule is connected to the smart meter and the power monitoring equipment to collect data item error calculation and module feedback.

[0016] By adopting the above technical solution, the historical load data of users is collected at certain time intervals through devices such as smart meters and stored in the database.

[0017] Preferably, the data of the time feature extraction submodule, the meteorological feature association submodule and the holiday feature identification submodule in the feature extraction module are processed by the data standardization submodule and the real-time data of the user.

[0018] The above technical solution is used to calculate the daily cycle characteristics and seasonal characteristics for processing.

[0019] Preferably, the model data of the model selection and construction submodule of the adaptive learning model module is provided synchronously by the data standardization submodule and the feature extraction module.

[0020] The above technical solution is adopted to train the model using the training data, and the parameters of the model are continuously optimized until the predetermined training goal or convergence condition is reached.

[0021] Preferably, the model architecture in the model selection and construction submodule is constructed using recurrent neural networks and their variants such as long short-term memory networks and gated recurrent units, and an integrated learning method is used to construct a strong learner.

[0022] By adopting the above technical solution, a deep neural network with multiple hidden layers can be selected to enhance the expression and learning capabilities of the model.

[0023] Preferably, the parameter optimization submodule is provided with algorithms such as Adam, Adagrad, and RMSProp to automatically adjust parameters such as learning rate and weight decay based on the gradient information of the training data.

[0024] The above technical solution enables the model to converge to the optimal solution more quickly during the training process.

[0025] Preferably, the model updating and adaptive adjustment submodule adopts a combination of online learning and incremental learning, and the model updating and adaptive adjustment submodule sets performance indicators in the monitoring model and prediction process.

[0026] By adopting the above technical solution, the model updating and adaptive adjustment submodule cooperates with the model selection and construction submodule to train the acquired historical data.

[0027] Preferably, the error calculation submodule uses error indicators to perform separate calculations and statistical analysis on the prediction errors in the prediction module.

[0028] By adopting the above technical solution, it is convenient to calculate and statistically analyze the prediction errors of different time scales separately through the error index.

[0029] Preferably, the feedback analysis submodule obtains the error calculation result of the error calculation submodule, performs feedback analysis on the prediction result and generates an error analysis report.

[0030] By adopting the above technical solution, the calculation result of the error calculation submodule is processed through the feedback analysis submodule.

[0031] Preferably, the error analysis report and collected data of the feedback analysis submodule are stored in a historical database, and the historical database provides data for model learning and training of the prediction module.

[0032] By adopting the above technical solution, through the adaptive learning processing of the historical database and the model, the model and parameters are continuously adjusted according to the real-time changes in data, which significantly improves the accuracy of the prediction.

[0033] Compared with the prior art, the beneficial effects of the present invention are: the user load forecasting system based on adaptive learning of historical load data:

[0034] 1. The data from smart meters and power monitoring equipment are collected and entered into the historical database through the data acquisition and preprocessing module. By establishing a data cache and update mechanism, the system can obtain the latest historical load data in real time. The collected data includes the user's timestamp, load value, voltage, current and other related electrical parameters. The adaptive learning model module selects appropriate deep learning and machine learning. Adaptive learning can better capture the law of load changes. The adaptive learning algorithm is used to automatically adjust the learning rate according to the changes in historical load data. The model is trained using training data and the model parameters are continuously optimized until the predetermined training target or convergence condition is reached. As the learning system deeply mines the complex patterns and potential laws in the historical load data, it can accurately predict the user load, thereby improving the accuracy of the prediction;

[0035] 2. The feature extraction module quickly differentiates and extracts weekly cycle features based on the historical composite data in the historical database, analyzes the load change rules of different working days and rest days in a week, and calculates the daily cycle features, divides a day into different time periods, and counts the load average, maximum, minimum, etc. of each time period. Taking into account seasonal characteristics, feature extraction based on the load characteristics of different seasons can adapt to the load characteristics and change trends of different users, and has strong versatility and flexibility;

[0036] 3. The error calculation submodule and feedback analysis submodule set up in the system jointly establish an error feedback channel, calculate the error between the predicted value and the actual load value in real time, find out the time period or prediction scenario with large error, and analyze the causes that may cause the error. At the same time, according to the results of feedback analysis, the model parameters, structure or position in the adaptive learning model module are automatically adjusted to reselect the model, which can adapt to the load characteristics and changing trends of various types of users. Accurate load forecasting helps to optimize the scheduling and resource allocation of the power system, reduce operating costs, and improve energy utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of the user load prediction system of the present invention;

[0038] Figure 2 This is a schematic diagram of the user load forecasting system module of the present invention;

[0039] Figure 3 It is a flow chart of the user load forecasting system of the present invention. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] See also Figure 1-Figure 3 ,The present invention provides a technical solution: a user load prediction system based on adaptive learning of historical load data, comprising a data acquisition and preprocessing module, a feature extraction module, an adaptive learning model module, a prediction module, an error evaluation and feedback module and a historical database;

[0042] The historical database is the main part of the user's electricity consumption historical load data record;

[0043] The data acquisition and preprocessing module is composed of a data acquisition submodule that is connected to the user's smart meter, power monitoring equipment and other data sources for communication and acquisition, a data cleaning submodule that performs quality inspection and processing on the collected data, and a data standardization submodule that performs standardization processing on the cleaned data and normalizes the load value and other data to a specific interval, and a data acquisition submodule that is connected to the smart meter and power monitoring equipment to collect data item error calculation and module feedback;

[0044] In conjunction with the accompanying drawings Figure 1-3As shown in the figure, the data sources in the data acquisition and preprocessing module obtain the user's smart meter, power monitoring equipment and other data sources through the data acquisition submodule, collect the user's historical load data in real time according to the preset time interval, and feed the acquired data into the historical database. By establishing a data cache and update mechanism, it is ensured that the system can obtain the latest historical load data in real time. At the same time, the historical data is classified and stored, for example, according to factors such as season, date type, weather conditions, etc., so that in the adaptive learning process, data patterns similar to the current prediction scenario can be quickly extracted and utilized. Data enhancement technology is used to appropriately transform the historical data, add noise, change the time scale of the data, etc., to increase the diversity of the data, so that the model can learn the load under different data changes. Prediction rules: After the setting of the data cleaning submodule, the collected data is quality tested and processed. Outliers in the data are detected, such as load values ​​that are obviously deviated from the normal range. They are identified and eliminated through statistical or clustering methods. Missing values ​​can be supplemented by linear interpolation, mean filling and other methods to ensure the integrity and continuity of the data. The cleaned data is standardized in conjunction with the data standardization submodule. Z-Score standardization or Min-Max standardization and other methods are used to make data of different magnitudes and units comparable. The load value and other data are normalized to a specific interval, such as [0,1] or [-1,1], to improve the efficiency and stability of model training, so that the data units remain consistent and the model is convenient for calculation and processing;

[0045] The feature extraction module is composed of a time feature extraction submodule that extracts the time-related features of the load data to construct a time series feature vector, a meteorological feature association submodule that collects meteorological data corresponding to the load data and analyzes the correlation between meteorological factors and load, and a holiday feature identification submodule that extracts and identifies holiday information in historical data. The data of the time feature extraction submodule, meteorological feature association submodule and holiday feature identification submodule in the feature extraction module are processed by the data standardization submodule to complete the user real-time data and the historical data from the historical database;

[0046] In conjunction with the accompanying drawings Figure 1-3As shown, the data processed by the data acquisition and preprocessing module is sent to the time feature extraction submodule, meteorological feature association submodule and holiday feature recognition submodule in the feature extraction module. The time feature extraction submodule constructs a time series feature vector by combining the input calendar information and the marked holiday mark. The daily load data is arranged into a vector in 24-hour time sequence as part of the model input so that the model can learn the time variation pattern of the load. At the same time, a day is divided into different time periods, such as morning, afternoon, and evening, and the load mean, peak value, and valley value of each time period are statistically obtained to obtain daily cycle characteristics. Then, the load variation rules of different working days and rest days in a week are analyzed to obtain weekly cycle characteristics. According to the load characteristics of different seasons in spring, summer, autumn, and winter, feature extraction and seasonal characteristics are performed to further differentiate data classification.

[0047] The meteorological characteristic association submodule collects meteorological data corresponding to load data, including temperature, humidity, air pressure, wind speed and other data, and analyzes the correlation between meteorological factors and load. It determines the meteorological characteristics that have a greater impact on the load through correlation analysis methods, quantifies them and integrates them into the characteristic vector, establishes a linear or nonlinear relationship model between temperature and load, and converts temperature characteristics into a form suitable for model input;

[0048] The holiday feature recognition submodule identifies holiday information in the historical database, extracts and identifies holidays as a special classification feature, and analyzes the difference between loads during holidays and non-holiday periods. The load change trend before and after holidays and the load fluctuation range during holidays provide the model with characteristic information about the impact of holidays on loads.

[0049] The adaptive learning model module is composed of a model selection and construction submodule that selects a suitable model architecture according to the characteristics of user load data and prediction requirements, a parameter optimization submodule that optimizes and adjusts the parameters of the model using an adaptive optimization algorithm, and a model update and adaptive adjustment submodule that monitors the performance indicators of the model in the prediction process in real time and adaptively adjusts them. The model data of the model selection and construction submodule of the adaptive learning model module is synchronously provided by the data standardization submodule and the feature extraction module. The model architecture in the model selection and construction submodule is constructed using a recurrent neural network and its variants such as a long short-term memory network and a gated recurrent unit, and a strong learner is constructed in conjunction with an integrated learning method. The parameter optimization submodule is provided with algorithms such as Adam, Adagrad, and RMSProp, and parameters such as the learning rate and weight decay are automatically adjusted based on the gradient information of the training data. The model update and adaptive adjustment submodule adopts a combination of online learning and incremental learning, and the model update and adaptive adjustment submodule sets performance indicators in the monitoring model and prediction process.

[0050] In conjunction with the accompanying drawings Figure 1-3 As shown in the figure, the model selection and construction submodule in the adaptive learning model module uses recurrent neural networks and their variants, long short-term memory networks and gated recurrent units. This architecture can handle long-term dependencies in time series data and adapt well to the time characteristics of load data. Through the unique gating mechanism of the long short-term memory network, it can remember the changing trend of load data over a long time span. When encountering similar time series patterns, it can better predict the user's future load data. Its learning data comes from the historical database, which adds the random forest or gradient boosting tree of the integrated learning method to the user's historical data. The combination constitutes multiple weak learners, which are further linked to construct a strong learner, which can adaptively The model can efficiently handle the complex relationship between different features and the noise in the data. The random forest reduces the overfitting risk of a single model by voting or averaging the results of multiple decision trees, so as to better adapt to various load data changes. The model uses Adam's adaptive learning rate algorithm to automatically adjust the learning rate according to the gradient history of each parameter. It calculates the first-order moment estimate and second-order moment estimate of the gradient, so that the learning rate can change dynamically according to the update frequency of the parameters and the gradient size. It automatically adjusts the learning rate, weight decay and other parameters according to the gradient information of the training data, so that the model can converge to the optimal solution faster during the training process. Regularization technology is applied during use to prevent model overfitting and improve the generalization ability of the model. By adjusting the regularization parameters, the complexity and fitting ability of the model are balanced. The model update and adaptive adjustment submodule adopts online learning and incremental learning to enable the model to be updated and adaptively adjusted in real time. When new load data arrives, the model can use these data to fine-tune the parameters to adapt to the real-time changes in load, thereby monitoring the performance indicators of the model during the prediction process. When the model performance declines or there are signs of overfitting, the model retraining or structural adjustment is automatically triggered to increase or decrease the number of hidden layer nodes, adjust the number of network layers, etc., to improve the prediction accuracy and adaptability of the model;

[0051] The prediction module is composed of a short-term prediction submodule for obtaining recent historical composite data for short-term prediction of user load, a medium-term prediction submodule for obtaining one month's historical load data for medium-term load prediction of the power system, and a long-term prediction submodule for long-term load prediction based on long-term trend analysis of historical data;

[0052] The error evaluation and feedback module is composed of an error calculation submodule for calculating the error between the predicted composite value and the actual load value of the prediction module, an error calculation submodule for performing feedback analysis on the prediction result based on the error calculation result of the error calculation submodule, and a model adjustment and optimization submodule for adjusting parameters, structures, and the favor selection model according to the results of the feedback analysis. The feedback analysis submodule obtains the error calculation result of the error calculation submodule, performs feedback analysis on the prediction result, and generates an error analysis report. The error analysis report of the feedback analysis submodule and the collected data are stored in the historical database, and the historical database provides data for the model learning and training of the prediction module. The error calculation submodule uses error indicators to perform separate calculations and statistical analysis on the prediction errors in the prediction module.

[0053] In conjunction with the accompanying drawings Figure 1-3As shown in the figure, the short-term prediction submodule in the prediction module makes a short-term prediction of the user's load in the next few hours to days, the medium-term prediction submodule makes a medium-term prediction of the user's load in the next week to month, and the long-term prediction submodule makes a long-term prediction of the user's load in the next few months to a year. The algorithm model in the prediction module adopts a method combining time series analysis and machine learning models, and uses the recent historical load data and related features for prediction. In the medium-term prediction process, the data model in the feature extraction comprehensively considers the periodic changes in historical loads, the medium-term trends of meteorological factors, and the seasonal changes in user behavior. Factors such as seasonal changes are jointly predicted, and a medium-term composite prediction report is provided. In the long-term prediction process, macro information such as economic development forecasts and policy factors are obtained through the Internet under the long-term trend characteristics of historical composite data. The macroeconomic model is combined with the power load model for prediction, and a long-term prediction report is produced. The above prediction report is fed back into the error evaluation and feedback module. At the same time, the error evaluation and feedback module obtains the user's real-time load data and compares it with the report data in the prediction module. The error calculation submodule uses commonly used error indicators to calculate the prediction load. The error between the load value and the actual load value is calculated and statistically analyzed for the prediction errors of different time scales, so as to more accurately evaluate the performance of the model in different prediction scenarios. Then, the data is sent to the feedback analysis submodule to find out the time period or prediction scenario with large errors, and analyze the possible causes of the errors. According to the error calculation results, the prediction results are fed back and analyzed to produce an error classification report, which provides a basis for the optimization and improvement of the model system in the adaptive learning model module. The model adjustment and optimization submodule pulls the report content of the feedback analysis submodule. When it is found that the load change law in a certain time period has changed significantly, resulting in an increase in the model prediction error, the error exceeds a certain threshold at this time, thereby automatically triggering the adjustment mechanism of the model adjustment and optimization submodule. By automatically adjusting the parameters and structure of the model or reselecting the model processing of the adaptive learning model module, the model is adjusted or retrained to better adapt to the new load mode. The data acquisition and preprocessing strategies are optimized to improve the data quality and the input accuracy of the model. At the same time, the feature extraction method is optimized, and relevant features are added or adjusted to enhance the model's sensitivity and prediction ability to load changes.

[0054] The adaptive learning process of this system provides key decision-making support for the operation and planning of the power system. Accurate load forecasting helps to optimize the dispatching strategy of the power system, rationally arrange power generation resources, avoid power shortages or surpluses, and thus reduce operating costs. Through more accurate load forecasting, the configuration of backup power generation capacity can be reduced, energy waste can be reduced, and energy utilization efficiency can be improved. At the same time, in terms of power system planning, it can provide a scientific basis for power grid construction and expansion, ensure that investment in power infrastructure matches actual demand, and promote the sustainable development of the power industry.

[0055] Working principle: When using the user load forecasting system based on adaptive learning of historical load data, firstly, the data sources such as user smart meters and power monitoring equipment are connected to the data acquisition submodule to communicate, and the user parameters are collected in real time according to the originally set time interval to form the latest historical load data and enter the historical database. After the data cleaning submodule uses the data cleaning algorithm to perform quality inspection and processing on the collected data, the cleaned data is standardized by the data standardization submodule to obtain standard data units. The time feature extraction submodule, meteorological feature association submodule and holiday feature recognition submodule set in the feature extraction module extract the periodic features of the load data, including daily cycle, The weekly and seasonal cycle characteristics are provided by the adaptive learning model module to provide a variety of deep learning and machine learning models. The learning rate is automatically adjusted according to the changes in historical load data. The model is trained using the training data in the historical database, and the model parameters are continuously optimized until the predetermined training target or convergence condition is reached. The prediction module predicts the load growth of the user's future composite data based on the user's historical load data, and the error evaluation and feedback module evaluates and corrects the error data to improve the model's sensitivity and prediction ability to load changes, providing strategic guidance for the long-term investment planning, grid construction and expansion decisions of the power system. The prediction results are presented in the form of annual load forecast curves and related statistical indicators.

[0056] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

Claims

1. A user load forecasting system based on adaptive learning of historical load data, characterized in that: include: It consists of data acquisition and preprocessing module, feature extraction module, adaptive learning model module, prediction module, error evaluation and feedback module and historical database; The data acquisition and preprocessing module is composed of a data acquisition submodule that collects data through communication with data sources such as smart meters and power monitoring equipment connected to users, a data cleaning submodule that performs quality inspection and processing on the collected data, and a data standardization submodule that performs standardization processing on the cleaned data and normalizes data such as load values ​​to a specific interval; The feature extraction module is composed of a time feature extraction submodule for extracting time-related features of load data to construct a time series feature vector, a meteorological feature association submodule for collecting meteorological data corresponding to load data and analyzing the correlation between meteorological factors and load, and a holiday feature identification submodule for extracting and identifying holiday information in historical data; The adaptive learning model module is composed of a model selection and construction submodule for selecting a suitable model architecture according to the characteristics of user load data and forecasting requirements, a parameter optimization submodule for optimizing and adjusting the parameters of the model using an adaptive optimization algorithm, and a model update and adaptive adjustment submodule for real-time update and adaptive adjustment of performance indicators of the monitoring model during the forecasting process; The prediction module is composed of a short-term prediction submodule for obtaining recent historical composite data for short-term prediction of user load, a medium-term prediction submodule for obtaining one month's historical load data for medium-term load prediction of the power system, and a long-term prediction submodule for long-term load prediction based on long-term trend analysis of historical data; The error evaluation and feedback module is composed of an error calculation submodule for calculating the error between the predicted composite value and the actual load value of the prediction module, an error calculation submodule for performing feedback analysis on the error calculation result of the error calculation submodule and the prediction result, and a model adjustment and optimization submodule for adjusting parameters, structure and favor selection model according to the results of the feedback analysis; The historical database is the main part of the user's electricity consumption historical load data record.

2. A user load forecasting system based on adaptive learning of historical load data according to claim 1, characterized in that: The data acquisition submodule is connected to the smart meter and the power monitoring equipment to collect data item error calculation and module feedback.

3. The user load forecasting system based on adaptive learning of historical load data according to claim 1 is characterized by: The data of the time feature extraction submodule, the meteorological feature association submodule and the holiday feature identification submodule in the feature extraction module are the user real-time data processed by the data standardization submodule and the historical data from the historical database.

4. The user load forecasting system based on adaptive learning of historical load data according to claim 1 is characterized by: The model data of the model selection and construction submodule of the adaptive learning model module is provided synchronously by the data standardization submodule and the feature extraction module.

5. The user load forecasting system based on adaptive learning of historical load data according to claim 1 is characterized by: The model architecture in the model selection and construction submodule is constructed using recurrent neural networks and their variants such as long short-term memory networks and gated recurrent units, and an integrated learning method is used to construct a strong learner.

6. The user load forecasting system based on adaptive learning of historical load data according to claim 1 is characterized by: The parameter optimization submodule is provided with algorithms such as Adam, Adagrad, and RMSProp to automatically adjust parameters such as learning rate and weight decay based on the gradient information of training data.

7. The user load forecasting system based on adaptive learning of historical load data according to claim 1 is characterized by: The model updating and adaptive adjustment submodule adopts a combination of online learning and incremental learning, and the model updating and adaptive adjustment submodule sets performance indicators in the monitoring model and prediction process.

8. The user load forecasting system based on adaptive learning of historical load data according to claim 1 is characterized by: The error calculation submodule uses error indicators to perform separate calculations and statistical analysis on the prediction errors in the prediction module.

9. The user load forecasting system based on adaptive learning of historical load data according to claim 1 is characterized by: The feedback analysis submodule obtains the error calculation result of the error calculation submodule, performs feedback analysis on the prediction result and generates an error analysis report.

10. The user load forecasting system based on adaptive learning of historical load data according to claim 1, characterized in that: The error analysis report and collected data of the feedback analysis submodule are stored in a historical database, and the historical database provides data for model learning and training of the prediction module.

Images