A method and system for predicting load of thermal power plant

Abstract

The present invention relates to the technical field of thermal power plants, and in particular to a method and system for predicting the load of a thermal power plant. The method comprises the following steps: obtaining real-time operation data of a thermal power plant, including historical load data, unit operation parameter data, environmental parameter data, and power grid dispatching data; performing data missing processing based on a time series interpolation method on the real-time operation data, and detecting and correcting abnormal values ​​to obtain operation preprocessing data; performing multi-scale time feature extraction on the operation preprocessing data to obtain time feature data, wherein the multi-scale time feature includes hourly features, daily features, and weekly features; and constructing a multi-level load prediction model based on the time feature data. The present invention comprehensively covers the load features at different time levels such as hours, days, and weeks through multi-scale time feature extraction, providing richer input information for prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal power plants, and in particular to a method and system for predicting the load of a thermal power plant. Background Art

[0002] A thermal power plant, also known as a coal-fired power plant, is a plant that uses combustibles (such as coal, oil, natural gas, etc.) as fuel to produce electricity. The working principles of a thermal power plant mainly include the following: chemical energy is converted into thermal energy (burning fossil fuels in a boiler to generate thermal energy, which is absorbed by water and becomes water vapor), thermal energy is converted into mechanical energy (the thrust generated by high-temperature steam drives the turbine to rotate), and mechanical energy is converted into electrical energy (using the rotation of the turbine to drive the generator rotor to rotate, and generating electrical energy by cutting the magnetic lines of force). The main systems of a thermal power plant include the fuel system, combustion system, steam-water system, electrical system, and control system. Among them, the most important equipment is the boiler, steam turbine, and generator, which are installed in the main plant of the power plant. The efficiency of a modern thermal power plant can reach 40%, that is, 40% of the thermal energy in the fuel is converted into electrical energy. The load of a thermal power plant refers to the amount of electricity that a thermal power plant needs to supply within a certain period of time. It is an important parameter for the dispatching of the power system and the operation and management of the thermal power plant. The prediction of the load of a thermal power plant is of great significance to the stable operation, economic dispatching, and energy conservation and emission reduction of the power system.

[0003] Traditional load forecasting methods often rely on historical load data and fail to fully consider real-time information such as unit operating parameters, environmental factors, and grid dispatch, resulting in low forecasting accuracy. Traditional models process load data mostly based on a single time scale, making it difficult to fully capture the load change characteristics within different time ranges. Summary of the invention

[0004] Based on this, it is necessary for the present invention to provide a method and system for predicting the load of a thermal power plant to solve at least one of the above technical problems.

[0005] To achieve the above object, a method for predicting the load of a thermal power plant comprises the following steps:

[0006] Step S1: obtaining real-time operation data of the thermal power plant, including historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data; performing data missing processing based on the time series interpolation method on the real-time operation data, and detecting and correcting abnormal values ​​to obtain operation preprocessing data;

[0007] Step S2: extracting multi-scale time features from the running preprocessing data to obtain time feature data, wherein the multi-scale time features include hourly features, daily features, and weekly features;

[0008] Step S3: construct a multi-level load forecasting model based on the time characteristic data, wherein the multi-level load forecasting model includes a short-term forecasting model based on random forest, a medium-term forecasting model based on XGBoost, and a long-term forecasting model based on long short-term memory network; train the multi-level load forecasting model using the running preprocessing data, and determine the model weight coefficient according to the characteristics of different forecast periods to obtain the load forecasting model;

[0009] Step S4: using the load prediction model to perform load prediction based on the pre-acquired current operating data to obtain initial prediction result data; performing sudden change operating condition detection based on the current operating data to obtain a sudden change operating condition feature vector, wherein the sudden change operating condition feature vector includes the output power change rate of the generator set, the boiler steam parameter change rate, the grid frequency change rate and the ambient temperature change rate;

[0010] Step S5: When the change rate of any parameter in the sudden change condition characteristic vector exceeds the preset threshold, the weight of the load prediction model is adaptively adjusted, and the prediction results of each model are weighted and fused according to the adjusted weight coefficient to obtain the load prediction data; when the change rate of the parameters used in the sudden change condition characteristic vector does not exceed the preset threshold, the initial prediction result data is used as the load prediction data.

[0011] The present invention can ensure the integrity and accuracy of data by acquiring real-time operation data including historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data, and using time series interpolation method to process missing data and correct outliers. This process provides reliable input data for subsequent load forecasting and improves the robustness and prediction accuracy of the prediction model. Multi-scale time feature extraction (hour, day, and week features) can capture the changing laws of power plant loads in different time dimensions. In this way, the model can better understand short-term, periodic and long-term trends, improve the adaptability to complex operating conditions, and ultimately improve the generalization ability and accuracy of the prediction model under different time spans. The use of a multi-level model based on random forest, XGBoost and long short-term memory network (LSTM) can give full play to the advantages of each algorithm. Random forest is suitable for noise filtering and nonlinear feature extraction in short-term prediction, XGBoost shows strong fitting ability in medium-term prediction, and LSTM is suitable for processing long-term trends. By training different models and determining the weight coefficients of the models according to the prediction period characteristics, this structure enables the model to have high prediction accuracy and stability at different time scales. By using the load forecasting model to forecast the pre-acquired current operating data, and combining it with the sudden change condition detection, the model can timely identify and respond to significant changes in the operating environment. The key parameters involved in the sudden change condition detection, such as the generator set output power, boiler steam parameters, grid frequency and ambient temperature, make the detection highly sensitive and comprehensive, which ensures that the model can respond quickly to emergencies and prevent the prediction results from being inaccurate under abnormal conditions. This step realizes the dynamic optimization of the model by adaptively adjusting the model weights. Specifically, when the change rate of the sudden change condition parameters exceeds the preset threshold, the model can adjust the weights in time to enhance the adaptability to the current operating environment. This adaptive mechanism improves the prediction accuracy of the model under sudden conditions and ensures the robustness and real-time performance of load prediction. In the case that there is no significant sudden change in the working condition, the initial prediction result is directly adopted, which reduces the consumption of computing resources and improves efficiency. Overall, the steps of this method complement each other, and through data preprocessing, feature extraction, model training, sudden change detection and weight adjustment, a complete closed loop from data input to load forecast output is formed. This process not only improves the accuracy and response speed of load forecasting, but also enhances the stability and practicality of the system under complex power plant operating conditions.

[0012] The present invention also provides a thermal power plant load prediction system, which is used to execute the thermal power plant load prediction method mentioned above. The thermal power plant load prediction system comprises:

[0013] The data acquisition and preprocessing module is used to obtain the real-time operation data of the thermal power plant, including historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data; the real-time operation data is processed for missing data based on the time series interpolation method, and abnormal value detection and correction are performed to obtain operation preprocessing data;

[0014] A time feature extraction module is used to extract multi-scale time features from the running preprocessing data to obtain time feature data, wherein the multi-scale time features include hourly features, daily features, and weekly features;

[0015] The prediction model building module is used to build a multi-level load prediction model based on time characteristic data, wherein the multi-level load prediction model includes a short-term prediction model based on random forest, a medium-term prediction model based on XGBoost, and a long-term prediction model based on long short-term memory network; the multi-level load prediction model is trained using running preprocessed data, and the model weight coefficient is determined according to the characteristics of different prediction periods to obtain the load prediction model;

[0016] The real-time load prediction module is used to use the load prediction model to perform load prediction based on the pre-acquired current operating data to obtain the initial prediction result data; perform sudden change condition detection based on the current operating data to obtain the sudden change condition feature vector, where the sudden change condition feature vector includes the output power change rate of the generator set, the boiler steam parameter change rate, the grid frequency change rate and the ambient temperature change rate;

[0017] The adaptive prediction adjustment module is used to adaptively adjust the weight of the load prediction model when the change rate of any parameter in the sudden change condition characteristic vector exceeds the preset threshold, and weightedly fuse the prediction results of each model according to the adjusted weight coefficient to obtain the load prediction data; when the change rate of the parameters used in the sudden change condition characteristic vector does not exceed the preset threshold, the initial prediction result data is used as the load prediction data.

[0018] The present invention collects real-time operation data of thermal power plants, including historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data, so that the system can fully understand the current working conditions and provide data support for subsequent load forecasting. The time series interpolation method is used to supplement the missing data, and outlier detection and correction are performed at the same time to ensure the accuracy and integrity of the data. The missing data is filled and the outliers are corrected by the interpolation method to ensure the continuity and consistency of the data, and avoid prediction errors caused by data problems; the cleaned data can improve the training and prediction accuracy of subsequent models, and avoid prediction errors or instability caused by data anomalies. Multi-scale time features are extracted from the operation preprocessing data, including hourly features, daily features and weekly features, to provide more detailed time dimension data for load forecasting. Through multi-scale feature extraction, the system can capture the load change rules on different time scales, which can reflect short-term fluctuations, daily cycle changes and long-term trends, etc., and improve the accuracy of the prediction model. The multi-scale extraction of time features enables the model to be analyzed at different time levels to better cope with different load change patterns. Construct a short-term prediction model based on random forest, a medium-term prediction model based on XGBoost, and a long-term prediction model based on long short-term memory network (LSTM). By training and running preprocessed data, a comprehensive load forecasting model is obtained by utilizing the characteristics and advantages of different models. Different models can provide optimization solutions for different forecasting periods. The short-term model adapts to the prediction of rapid fluctuations, the medium-term model provides a relatively stable trend prediction, and the long-term model can capture longer-term trend changes to form a hierarchical prediction capability. By combining different prediction algorithms such as random forest, XGBoost and LSTM, the system can better adapt to load fluctuations at different time scales and improve the accuracy of overall load forecasting. The system can quickly perform load forecasting with the support of real-time data to meet the real-time prediction needs of power plants for load demand; by mutating the operating condition feature vectors (such as power, steam, frequency and temperature change rate), it can identify in real time whether there are abnormal operating conditions, adjust the load forecasting strategy in time, and prevent errors under sudden operating conditions. Under sudden operating conditions, by adaptively adjusting the weights of each sub-model, the prediction deviation caused by sudden operating conditions can be avoided to ensure the accuracy of load forecasting data. This module can dynamically adjust the prediction model under different operating conditions, making the system flexible and robust when facing sudden operating conditions. When the system is running normally, the initial prediction results are directly used to avoid excessive adjustments or unnecessary complex calculations, saving computing resources and ensuring efficiency; under sudden operating conditions, dynamic adjustments are made to make the load prediction results more accurate and avoid errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0020] Figure 1 It is a schematic flow chart of the steps of the method for predicting the load of a thermal power plant of the present invention;

[0021] Figure 2 for Figure 1 Detailed step flow diagram of step S1;

[0022] Figure 3 for Figure 1 Detailed step flow chart of step S2 in FIG. DETAILED DESCRIPTION

[0023] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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 technicians in this field without creative work are within the scope of protection of the present invention.

[0024] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0025] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0026] To achieve this, please refer to Figures 1 to 3 The present invention provides a method for predicting the load of a thermal power plant, the method comprising the following steps:

[0027] Step S1: obtaining real-time operation data of the thermal power plant, including historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data; performing data missing processing based on the time series interpolation method on the real-time operation data, and detecting and correcting abnormal values ​​to obtain operation preprocessing data;

[0028] The embodiment of the present invention obtains real-time operation data of a thermal power plant, specifically including historical load data extracted from a database, boiler steam parameters, turbine parameters and overall output power data of a generator set collected through a unit monitoring system as unit operation parameter data, and also obtains environmental parameter data such as ambient temperature, humidity and air pressure through a meteorological station, and obtains power grid dispatch data such as planned power generation and power grid frequency from a power grid dispatch center. First, linear interpolation is used to fill in missing data less than 1 hour, and cubic spline interpolation is used to fill in missing data greater than 1 hour, to complete data missing processing. Then, the Z-score method and isolation forest algorithm are used to detect outliers for the complete operation data, and a range check of the physical constraints of the equipment operation is performed on the data points marked as abnormal, and the moving median replacement method is performed to correct the smoothed operation preprocessing data.

[0029] Step S2: extracting multi-scale time features from the running preprocessing data to obtain time feature data, wherein the multi-scale time features include hourly features, daily features, and weekly features;

[0030] The embodiment of the present invention performs multi-scale time feature extraction on the operation preprocessing data generated in step S1, and extracts time features at hourly, daily and weekly scales. In the hourly scale feature extraction, the average load, load change rate, peak-to-valley difference and load fluctuation index per hour are calculated; in the daily scale feature extraction, the daily maximum load, minimum load, daily load rate, daily peak-to-valley difference and shape characteristics of the daily load curve are analyzed; in the weekly scale feature extraction, the weekly load change pattern, weekly load distribution characteristics and weekly cycle pattern are identified. After the feature data of different time scales are subjected to correlation analysis to identify the correlation between the features, the principal component analysis method is used to reduce the dimensionality of the features to obtain the final time feature data.

[0031] Step S3: construct a multi-level load forecasting model based on the time characteristic data, wherein the multi-level load forecasting model includes a short-term forecasting model based on random forest, a medium-term forecasting model based on XGBoost, and a long-term forecasting model based on long short-term memory network; train the multi-level load forecasting model using the running preprocessing data, and determine the model weight coefficient according to the characteristics of different forecast periods to obtain the load forecasting model;

[0032] According to the time characteristic data generated in step S2, the embodiment of the present invention constructs short-term, medium-term and long-term load forecasting sub-models respectively, wherein the short-term forecasting model is constructed based on the random forest algorithm, the medium-term forecasting model is constructed based on the XGBoost algorithm, and the long-term forecasting model is constructed using the long short-term memory network algorithm (LSTM). The running preprocessing data is divided into a training set, a validation set and a test set in a ratio of 8:1:1 and standardized, and then the hyperparameters of each sub-model are optimized using the grid search method. The specific parameter range includes the number of trees of the random forest of 100-500, the learning rate of XGBoost of 0.01-0.3, the number of hidden layers of LSTM of 1-3, etc. The weights of each model on different time scale features are evaluated according to the prediction accuracy of the validation set, and finally a load forecasting model is generated.

[0033] Step S4: using the load prediction model to perform load prediction based on the pre-acquired current operating data to obtain initial prediction result data; performing sudden change operating condition detection based on the current operating data to obtain a sudden change operating condition feature vector, wherein the sudden change operating condition feature vector includes the output power change rate of the generator set, the boiler steam parameter change rate, the grid frequency change rate and the ambient temperature change rate;

[0034] The embodiment of the present invention performs prediction based on the currently acquired operating data and obtains initial prediction result data by calling the load prediction model obtained in step S3. In the load prediction process, the short-term, medium-term and long-term models are respectively called, and the prediction data are integrated through the model fusion weight. Subsequently, the sudden change condition detection is performed based on the current operating data, and the output power change rate of the generator set, the boiler steam parameter change rate, the grid frequency change rate and the ambient temperature change rate are calculated to obtain the sudden change condition feature vector, which is used to determine whether the weight of the prediction model needs to be adjusted.

[0035] Step S5: When the change rate of any parameter in the sudden change condition characteristic vector exceeds the preset threshold, the weight of the load prediction model is adaptively adjusted, and the prediction results of each model are weighted and fused according to the adjusted weight coefficient to obtain the load prediction data; when the change rate of the parameters used in the sudden change condition characteristic vector does not exceed the preset threshold, the initial prediction result data is used as the load prediction data.

[0036] The embodiment of the present invention compares the parameter change rate in the sudden working condition characteristic vector obtained in step S4 with the preset threshold. When the change rate of any parameter in the sudden working condition characteristic vector exceeds the set threshold, the weight adaptive adjustment mechanism is started, and the weights of each sub-model in the load forecasting model are recalculated based on the sudden working condition characteristics to obtain optimized weight data, and then the adjusted weight coefficient is used to weightedly fuse the results of each prediction sub-model to obtain the final load forecast data; if all parameter change rates are within the preset threshold, the initial prediction result data is directly output as the final load forecast data.

[0037] Preferably, step S1 comprises the following steps:

[0038] Step S11: extracting historical power generation load curves and data points from the thermal power plant database to obtain historical load data;

[0039] The embodiment of the present invention uses the interface query module of the thermal power plant database to call the data extraction script to extract the load curve data and related data points from the historical power generation load record table. First, the time range and frequency of data extraction are set, and the hourly or minute-level load data of the past two years are usually selected to cover the long-term and short-term load changes of the power plant. During the extraction process, the data points are time-marked using the timestamp alignment algorithm to ensure that each data point matches the power generation load at the corresponding time, and finally the historical load data arranged in time series is obtained, which is suitable for subsequent analysis and prediction.

[0040] Step S12: Collecting boiler steam parameters, turbine parameters and overall output power of the generator set in the unit operation monitoring system, thereby obtaining unit operation parameter data;

[0041] The embodiment of the present invention connects the real-time data acquisition module through the unit operation monitoring system to obtain parameters such as boiler steam temperature, steam pressure, turbine speed, turbine power, and generator output power. These parameters are recorded by timestamp to form a monitoring data stream under the unit operation state, and the parameters of each unit are continuously collected with a sampling interval of 1 minute. The data cleaning module is used to remove interference data and duplicate data, and finally stable unit operation parameter data is obtained to ensure its accuracy to support subsequent load forecasting analysis.

[0042] Step S13: Obtain environmental parameter data through a weather station, including ambient temperature, humidity and air pressure data;

[0043] The embodiment of the present invention obtains environmental monitoring data in real time, including ambient temperature, relative humidity and air pressure information, by accessing the meteorological station database of the thermal power plant. The average value calculation function built into the meteorological station is used to average the collected data within 1 hour to ensure the representativeness of the data. At the same time, the meteorological data is sorted by time label and imported into the real-time data table for subsequent use in combination with other operating data. The data can be used to analyze the impact of environmental factors on power generation load, which helps to improve the accuracy of the load forecasting model.

[0044] Step S14: acquiring power grid dispatching data through the power grid dispatching center, including planned power generation data and power grid frequency data;

[0045] The embodiment of the present invention calls the power generation dispatch data acquisition interface through the data interface of the power grid dispatch center to extract real-time data including dispatch information such as planned power generation and power grid frequency. The query frequency is set to every hour, and the data is time-aligned to ensure time consistency with the thermal power plant operation data. For periods of large power grid frequency fluctuations, the data collection frequency is increased to capture details, and finally complete power grid dispatch data is obtained for input features of subsequent load forecasting models.

[0046] Step S15: merging historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data into real-time operation data;

[0047] The embodiment of the present invention integrates historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data into unified real-time operation data. First, the data from different sources are aligned by timestamp, and then the data fusion module is used to splice various types of data into an integrated data table in time series. In the data fusion process, multiple data at the same time point are combined and calculated, such as multi-index weighted processing of environmental data and unit parameter data, to obtain more accurate operation data, and finally form a real-time operation data set for application in load forecasting analysis.

[0048] Step S16: Perform data missing processing based on the time series interpolation method on the real-time operation data, and detect and correct abnormal values ​​to obtain operation preprocessing data.

[0049] The embodiment of the present invention performs data missing processing on the real-time operation data generated in step S15, and fills it with time series interpolation. Specifically, linear interpolation is used to fill in the data with a missing duration of less than 1 hour, and cubic spline interpolation is used to fill in the data with a missing duration of more than 1 hour to ensure the continuity of the data. Subsequently, the isolation forest algorithm is used to detect outliers, and the abnormal data is marked and replaced with a moving median. The processed data is recorded as abnormal operation preprocessing data after re-detection, providing a basis for subsequent load forecasting.

[0050] The present invention extracts historical power generation load curves and data points from the thermal power plant database, which can provide a detailed historical data basis for subsequent load forecasting. This data can reflect the change trend and fluctuation characteristics of the load over time, help the prediction model to identify long-term and periodic patterns, and improve the accuracy and stability of the model in long-term load forecasting. Boiler steam parameters, turbine parameters and the overall output power of the generator set are collected from the unit operation monitoring system to ensure comprehensive monitoring of the unit operation status. These data can reflect the response of the power generation equipment under different load conditions, help to improve the detail accuracy of the model, and make the prediction results more reflect the actual operating characteristics of the unit. Environmental parameter data (such as temperature, humidity, and air pressure) obtained from the meteorological station can help the model understand the impact of the external environment on the power generation load. These factors often have a significant effect on power generation and power demand. For example, ambient temperature can affect the efficiency of the cooling water system and the overall power generation load demand. Including these data helps to improve the accuracy and adaptability of load forecasting. Grid dispatching data (such as planned power generation and grid frequency) provides the model with comprehensive information on the power supply and demand sides. Planned generation data enables forecasts to be consistent with actual dispatch plans, while grid frequency data reflects load balance and helps identify system anomalies. The introduction of these data supports real-time forecasting of load demand and prevention of potential dispatch conflicts. The integration of historical load data, unit operating parameters, environmental parameters and grid dispatch data into real-time operation data provides a complete input data set, which helps to improve the availability of data and the comprehensive performance of the forecasting model. Through this merging method, the model can more comprehensively understand the complex interaction between the thermal power plant and its external environment. The use of time series interpolation method to handle missing data and perform outlier detection and correction can greatly improve the integrity and quality of data. High-quality data input is an important prerequisite for the reliability of the load forecasting model, preventing forecast errors caused by missing or abnormal data. Through this step, the continuity and consistency of the data are ensured, thereby enhancing the accuracy and stability of the model. Through these steps, the method realizes the complete process from multi-data source acquisition to data preprocessing, ensuring the breadth, depth and accuracy of the data. This process provides a solid data foundation for the load forecasting model, helps to improve the accuracy and real-time performance of load forecasting in the operation of thermal power plants, and improves the model's adaptability to complex working conditions.

[0051] Preferably, step S16 comprises the following steps:

[0052] Step S161: performing data missing detection on the real-time operation data and performing data missing processing to obtain complete operation data, wherein the data missing processing specifically uses linear interpolation to process missing data less than 1 hour, and uses cubic spline interpolation to process missing data greater than 1 hour;

[0053] The embodiment of the present invention uses a data integrity check module to perform missing value detection on each data point in the real-time operation data, and records the time point and duration of the missing data. For the case where the missing duration is less than 1 hour, linear interpolation is used for processing, and linear interpolation calculation is performed using the data points before and after the missing value to fill the missing value. For the case where the missing duration exceeds 1 hour, cubic spline interpolation is used to obtain a smooth curve by fitting multiple data points before and after, and fill the missing values ​​according to the curve interpolation to ensure data continuity. After the filling is completed, complete operation data without missing data is obtained to support subsequent anomaly detection.

[0054] Step S162: performing outlier detection based on the Z-score method and the isolation forest algorithm on the complete operation data, and marking the abnormal data points at the identified locations, thereby obtaining outlier marked data, wherein the outlier marked data includes compound marked abnormal data and single marked abnormal data;

[0055] The embodiment of the present invention applies the Z-score method to the complete operation data. First, the mean and standard deviation of the data are calculated, each data point is standardized to a Z-score value, ±3 is set as the abnormal threshold, and abnormal data points exceeding the threshold are marked. Then, the isolation forest algorithm is used to further detect the data, and abnormal data points are identified based on feature isolation, and data with abnormal scores higher than the set threshold are marked as abnormal. Finally, the marked outlier data is obtained, which contains composite marked abnormal data (satisfying both Z-score and isolation forest conditions) and single marked abnormal data (satisfying only one condition), providing a basis for subsequent processing.

[0056] Step S163: Evaluate the degree of data deviation between the single marked abnormal data and the previous and next time windows, and perform a range check based on the physical constraints of the equipment operation, so as to obtain marked abnormal detection data, wherein the previous and next time windows are specifically ±15 minutes;

[0057] For single-marked abnormal data, the embodiment of the present invention extracts the data window of ±15 minutes before and after to evaluate the degree of deviation, calculates the average deviation between the current abnormal data and the data in the window, and determines whether it is within the physical constraint range of the equipment (such as the upper and lower limits of boiler steam temperature, generator power range, etc.). When the deviation value exceeds the physical constraint range, it is confirmed as a possible abnormal value and marked as marked abnormal detection data for further processing to ensure that the detection data can accurately reflect the actual operation of the equipment and provide a reference for subsequent screening.

[0058] Step S164: screening the single labeled abnormal data for abnormal data points according to the labeled abnormal detection data, thereby obtaining abnormal value labeled screening data;

[0059] The embodiment of the present invention further screens single marked abnormal data points based on marked abnormal detection data. First, the abnormal points with a large degree of deviation are recorded as abnormal value marked screening data and marked as high priority processing objects. Then, for abnormal points that do not meet the physical constraint range, they are directly confirmed as screened abnormal data. This screening step can effectively filter non-sudden anomalies, improve the accuracy of subsequent data correction, ensure that most of the noise data has been removed before smoothing, and improve the reliability of load forecasting.

[0060] Step S165: merging the composite-marked abnormal data and the abnormal value marked screening data into comprehensive abnormal value data;

[0061] The embodiment of the present invention combines the composite-marked abnormal data and the abnormal value marked screening data into comprehensive abnormal value data. First, the two types of abnormal data are merged by timestamp to avoid duplicate marking. For multiple marked abnormal points at the same time, the abnormal information of the composite mark is recorded first to ensure the integrity and consistency of the comprehensive abnormal value data. The combined comprehensive abnormal value data will be used as a basis in the subsequent correction process to avoid error propagation caused by missing abnormal data.

[0062] Step S166: According to the comprehensive abnormal value data, the complete operation data is corrected based on the moving median replacement method, and data smoothing is performed to obtain operation preprocessing data.

[0063] The embodiment of the present invention uses the moving median replacement method to correct the abnormal points in the complete operation data based on the comprehensive abnormal value data. The specific operation is: the median is calculated within the time window of each abnormal point (usually ±5 data points), and the median is used to replace the abnormal value, thereby reducing the impact of the abnormality on the data. Subsequently, the exponential smoothing method is applied to smooth the corrected data to eliminate the fluctuations caused by data mutations. Finally, the smoothed operation preprocessing data is obtained to provide stable and abnormal-free input data for the load forecasting model.

[0064] The real-time operation data in the present invention is often limited by the collection equipment or transmission problems and data loss occurs. By processing missing data less than 1 hour by linear interpolation, data gaps in short time intervals can be quickly and effectively filled to ensure data continuity. For missing data greater than 1 hour, the cubic spline interpolation method is used to better maintain the smoothness and trend of the data curve, which is suitable for processing data segments missing for a longer time, ensuring the rationality and accuracy of data filling. Using the Z-score method and the isolation forest algorithm to detect outliers, global and local features can be taken into account, and values ​​and complex abnormal patterns that are statistically significantly deviated from the normal range can be identified. The Z-score method is simple and efficient, suitable for detecting single variable anomalies, while the isolation forest algorithm can identify anomalies in high-dimensional data. By marking composite and single abnormal points, it helps to distinguish different degrees of anomalies and improve the meticulousness and reliability of detection. By evaluating the degree of data deviation in the time window (±15 minutes) before and after the single-labeled abnormal data, it can be identified whether there is a non-significant anomaly caused by short-term fluctuations. This, combined with the physical constraints of the equipment's operation, helps to ensure that data anomalies are caused by real equipment problems, rather than measurement or transmission errors, and improves the accuracy of anomaly detection. Screening single-labeled abnormal data based on labeled abnormal detection data helps reduce the amount of data that is misjudged or unnecessarily corrected, ensuring that only truly abnormal data is processed. This step of controlling accuracy helps improve the efficiency of data correction and the accuracy of subsequent prediction models. Combining composite labeled abnormal data and filtered single-labeled abnormal data into comprehensive outlier data helps to fully understand data anomalies and integrate multi-source detection results. This ensures the comprehensiveness of anomaly detection data and facilitates subsequent unified correction processing, thereby avoiding missing or repeated processing of abnormal data points. Correcting data based on the moving median replacement method can reduce the impact of outliers on the overall data while maintaining the stability and continuity of the data. The median method better resists the impact of extreme outliers, while data smoothing reduces the fluctuation of the corrected data, making the operating data more stable and consistent with physical meaning, providing high-quality data input for subsequent analysis and model training, and thus improving the accuracy and stability of the load forecasting model. These steps form a systematic preprocessing method from data missing detection, outlier detection to data correction and smoothing. This method ensures the integrity, accuracy and stability of real-time operation data, greatly reduces the interference of data problems on the load forecasting model, and improves the overall performance and credibility of the forecasting model.

[0065] Preferably, step S2 comprises the following steps:

[0066] Step S21: extracting hourly features of the load curve in the hourly time window of the operation preprocessing data, thereby obtaining hourly feature data, wherein the hourly features include hourly average load, load change rate, peak-to-valley difference, and load fluctuation index;

[0067] The embodiment of the present invention divides the operation preprocessing data into time windows by hour, and calculates the hourly characteristics of the load curve data for each hour separately. Specifically, the average load value is calculated for each hour of data as the average load per hour, reflecting the average level of the load in this time period; at the same time, the change rate between the starting load and the ending load in this hour is calculated and recorded as the load change rate to analyze the load fluctuation trend; in addition, the highest load and the lowest load in each hour are identified, and the difference between the two is calculated to obtain the peak-to-valley difference to reflect the fluctuation range of the hourly load; finally, the load fluctuation index is obtained by calculating the variance of the load data, and the stability of the load on the hourly scale is evaluated, and finally the hourly characteristic data is formed to provide detailed support for the subsequent time characteristic analysis.

[0068] Step S22: extracting features from the operation preprocessing data based on the daily cycle, thereby obtaining daily feature data, wherein the daily feature data includes daily maximum load, daily minimum load, daily load rate, daily peak-to-valley difference, and daily load curve shape features;

[0069] The embodiment of the present invention extracts features from the operation preprocessing data according to the daily cycle, and extracts the main load features from the daily data to evaluate the distribution and changes of the load in the daily cycle. The specific operation includes first calculating the daily maximum load and minimum load, recording them as the daily maximum load and the daily minimum load, which represent the daily peak and trough levels respectively; then calculating the daily load rate, that is, the ratio of the daily average load to the maximum load, which is used to evaluate the utilization rate of the load operation; at the same time, calculating the daily peak-to-valley difference to reflect the fluctuation range of the daily load; finally, based on the morphological analysis of the load curve, the shape characteristics of the daily load curve are obtained, and the curve fitting or morphological distribution feature extraction method (such as polynomial fitting) is used to describe the changing trend of the load curve, and the daily characteristic data is obtained to provide periodic support for the prediction model.

[0070] Step S23: Perform weekly scale feature analysis on the operation preprocessing data to extract weekly load variation rules, weekly load distribution characteristics and weekly cycle characteristics, thereby obtaining weekly feature data;

[0071] The embodiment of the present invention groups the operation preprocessing data according to the weekly cycle to extract the load variation law on the weekly scale. Specifically, firstly, the variation trend of the load within the week is analyzed, and the difference in load distribution between working days and non-working days is extracted and recorded as the weekly load variation law to reflect the basic fluctuation mode of the load on the weekly scale; then, the average load, maximum load and minimum load of each day of the week are calculated, and their distribution characteristics are analyzed and recorded as the load distribution characteristics within the week to identify the peak or trough date; finally, the repeatability characteristics of the load within the week are identified, and the weekly cycle characteristics are determined by calculating the correlation of the adjacent weekly data, and the weekly characteristic data is obtained to ensure that the model has high accuracy and reliability on the weekly time scale.

[0072] Step S24: performing correlation analysis on the hourly feature data, the daily feature data, and the weekly feature data to identify the correlation relationship between features of different time scales, thereby obtaining feature correlation data;

[0073] The embodiment of the present invention performs correlation analysis between the extracted hourly feature data, daily feature data, and weekly feature data to identify the correlation between features of different time scales. The specific operation is as follows: first, the Pearson correlation coefficient method is used to calculate the correlation between each time scale feature to identify highly correlated feature pairs; partial correlation analysis is further used for significantly correlated features to eliminate pseudo-correlated features affected by other factors; finally, feature correlation data is obtained, including significantly correlated feature groups and potential causal relationship features, which provide a basis for model feature selection and combination and reduce the interference of redundant information on prediction.

[0074] Step S25: sorting the importance of each time scale feature according to the feature association data, and performing feature dimension reduction and combination based on the principal component analysis method, so as to obtain time feature data.

[0075] The embodiment of the present invention uses feature association data to sort the priority of each time scale feature, and performs feature dimension reduction and combination based on principal component analysis (PCA). First, the importance of each time scale feature is determined by sorting, and the features with high importance are retained first. Subsequently, PCA is applied to combine highly correlated features into principal component features, and the representative principal components are reduced in dimension, thereby compressing redundant feature data. The final output time feature data will only retain the most valuable features for load forecasting, ensuring the computational efficiency and accuracy of the prediction model during training.

[0076] The present invention can accurately describe the basic trend and fluctuation characteristics of hourly load by extracting features (such as hourly average load, load change rate, peak-to-valley difference and load fluctuation index) from hourly load data. These hourly features help capture the short-term changes in daily load, provide accurate data reflecting dynamic changes in load for the load forecasting model, and are particularly helpful for short-term load forecasting. The hourly average load can reflect the overall level of load during the period. The load change rate can show the fluctuation amplitude of load within the hour. The peak-to-valley difference and the load fluctuation index help evaluate the volatility and fluctuation trend of load, so as to discover potential load anomalies and system stability problems. Periodic analysis of daily load data (such as daily maximum load, daily minimum load, daily load rate, daily peak-to-valley difference and daily load curve shape characteristics) helps to reveal the law of daily load fluctuations, identify the peak period, trough period and overall pattern of load fluctuations of daily load. These features are particularly important for medium-term load forecasting, which can help the model capture the intraday periodic law and accurately predict similar fluctuations that may occur in the next few days. The daily maximum load and daily minimum load reflect the daily load fluctuation range, which helps to discover the load pressure in different time periods. Daily load rate and daily peak-to-valley difference help evaluate the distribution and fluctuation of load within a day. The shape characteristics of daily load curve further reveal the trend and pattern of load change, which helps to analyze the long-term fluctuation characteristics of load. By extracting features from weekly data (such as weekly load change law, weekly load distribution characteristics, and weekly cycle characteristics), the periodic changes of load in different weeks can be captured, so as to understand the load fluctuation and trend between different weeks. This is particularly critical for long-term prediction models, especially for revealing the relationship between periodic load fluctuations, seasonal changes, and influencing factors. Weekly load change law helps identify the trend of weekly load fluctuations and find out the load differences between different working days. Weekly load distribution characteristics help analyze the changes in the load difference between working days and rest days. Weekly cycle characteristics reveal the periodic law of load in different time periods within the week, which helps to predict the long-term load fluctuation trend. By performing correlation analysis on hourly features, daily features, and weekly features, the correlation between features of different time scales can be identified. This analysis helps to understand the inherent connection between load changes in different time periods, reveal the relationship between short-term load fluctuations and long-term load trends, and provide a basis for subsequent feature selection, model training, and result optimization. Feature correlation data optimizes the input features of the model by revealing the interactions between different time scales, allowing the model to use these important correlations for more accurate predictions. By sorting the importance of features, the most influential features for load forecasting can be selected, redundant features can be reduced, and the prediction performance and computational efficiency of the model can be improved. In addition, principal component analysis (PCA) can reduce the dimensionality of data and reduce the number of features while retaining the most important change information in the data.This can reduce the computational burden brought by high-dimensional data, reduce the risk of overfitting, and improve the generalization ability of the model. Feature dimensionality reduction and combination remove irrelevant or redundant features by compressing data dimensions, allowing the model to focus on the information that contributes most to the prediction, thereby improving model efficiency and accuracy. These steps help build more accurate and efficient load forecasting models by extracting and optimizing load features from different time scales (hours, days, weeks). These steps ensure the high quality and richness of feature data, enabling the model to capture the laws of load changes from multiple time dimensions, ultimately improving the accuracy, reliability and computational efficiency of load forecasting.

[0077] Preferably, step S3 comprises the following steps:

[0078] Step S31: construct short-term, medium-term and long-term load forecasting sub-models based on time characteristic data, wherein the short-term forecast adopts the random forest algorithm, the medium-term forecast adopts the XGBoost algorithm, and the long-term forecast adopts the long short-term memory network algorithm, thereby obtaining a multi-level load forecasting model;

[0079] The embodiment of the present invention constructs short-term, medium-term and long-term load forecasting sub-models according to time characteristic data. First, the random forest algorithm is used to construct a short-term prediction sub-model. The algorithm improves the prediction accuracy by constructing multiple decision trees and taking their average prediction value, which is suitable for hourly load forecasting; secondly, the XGBoost algorithm is used to construct a medium-term prediction sub-model. The algorithm is based on the gradient boosting decision tree and has good generalization ability, which is suitable for daily load forecasting; finally, the long-term short-term memory network (LSTM) is used to construct a long-term prediction sub-model. The network can remember long-term dependent features and is suitable for weekly scale forecasting. By using adaptive algorithms for different time scales to construct a multi-level load forecasting model, support is provided for improving the accuracy of load forecasting.

[0080] Step S32: Divide the running preprocessed data into a training set, a validation set, and a test set in a ratio of 8:1:1, and standardize the data to obtain model training data;

[0081] The embodiment of the present invention divides the running preprocessing data into a training set, a validation set, and a test set in a ratio of 8:1:1 to ensure the representativeness and balance of the data; the training set is used for model training, the validation set is used for model parameter adjustment and optimization, and the test set is used for the final performance evaluation of the model. Then, the divided data set is standardized, and each feature value is converted to the same range (for example, 0 to 1 or -1 to 1) to eliminate the impact of data scale differences on model training, thereby obtaining homogenized model training data to ensure stability during model training and accuracy of prediction.

[0082] Step S33: using a grid search method to optimize the hyperparameters of each prediction sub-model in the multi-level load prediction model according to the training set in the model training data, thereby obtaining an optimized prediction model;

[0083] The embodiment of the present invention uses a grid search method to optimize the hyperparameters of each sub-model in a multi-level load forecasting model. The specific operation is: first define the hyperparameter search space of each sub-model, such as the number of trees and maximum depth of random forest, the learning rate and maximum depth of XGBoost, the number of layers and hidden units of LSTM, etc.; then, use the grid search method to traverse these parameter combinations in turn, and train them on the training set, and calculate the prediction accuracy of each group of parameters on the validation set. Finally, the parameter combination with the best prediction effect is selected as the final hyperparameter of the model to obtain an optimized prediction model, thereby improving the prediction accuracy of the model at different time scales.

[0084] Step S34: evaluating the importance of the features of each prediction period in the optimized prediction model, calculating the influence of features of different time scales on the prediction results, and thus obtaining feature weight data;

[0085] The embodiment of the present invention evaluates the importance of each prediction period feature in the optimized prediction model to analyze the impact of different time scale features on the prediction results. The specific approach is: by calculating the contribution of each feature in the model prediction, the importance ranking of the features is obtained. For example, random forest and XGBoost can evaluate the importance of the feature by the number of splits and split gains in the tree, and LSTM can analyze the feature weights through the attention mechanism or back propagation gradient. Finally, the weight value of each feature in the model is obtained to form feature weight data, which is used to guide the model's optimal feature selection in different prediction periods.

[0086] Step S35: Based on the validation set in the model training data, the prediction accuracy of each sub-model of the optimized prediction model at different prediction time scales is evaluated to obtain model evaluation data;

[0087] Based on the optimization of the prediction model, the embodiment of the present invention uses the validation set data to evaluate the prediction accuracy of each sub-model at different prediction time scales. The specific operation is: predict each sub-model at the hour, day and week scales, calculate the corresponding error indicators (such as mean square error, mean absolute error, etc.), and record its prediction accuracy. The model evaluation data obtained in this step will be used for the adaptive weight allocation of subsequent models to improve the adaptability of the model at different time scales and ensure the final load prediction accuracy.

[0088] Step S36: Determine the weight coefficients of each sub-model in different prediction periods based on the adaptive weight allocation method according to the model evaluation data and the feature weight data, thereby obtaining model weight data;

[0089] The embodiment of the present invention uses an adaptive weight allocation method to determine the weight coefficient of each sub-model in different prediction periods based on the model evaluation data and feature weight data. The specific method is: combining the importance weights of the features of each time scale and the model evaluation results, allocating the optimal weights at different time scales through an optimization algorithm (such as weighted linear regression), and generating model weight data. The weight data can flexibly adjust the contribution of each sub-model in the prediction, so that the model has a high adaptability in both sudden load and stable load scenarios.

[0090] Step S37: Structural optimization is performed on the optimization prediction model according to the model weight data, thereby obtaining a load prediction model.

[0091] The embodiment of the present invention further optimizes the structure of the optimized prediction model according to the model weight data to generate the final load prediction model. The specific operation is: combining the weight data, adjusting the weight coefficients of each sub-model and retraining the model structure so that it can adapt to the different time scale characteristics and dynamic change characteristics of the load. The load prediction model finally obtained is more optimized in structure, has good generalization ability and adaptability, and can accurately predict the load of thermal power plants.

[0092] The present invention can select the most appropriate prediction method for the load change characteristics of different time scales by respectively adopting algorithms suitable for prediction at different time scales (random forest for short-term prediction, XGBoost for medium-term prediction, and long short-term memory network (LSTM) for long-term prediction), thereby obtaining an optimized model for each prediction period. Short-term prediction: Random forest can process high-dimensional data and can capture the nonlinear characteristics of short-term load fluctuations, which is very suitable for capturing hourly load changes. Medium-term prediction: XGBoost has advantages in processing features with strong correlations and solving overfitting problems, and is suitable for load prediction on a daily or weekly scale. Long-term prediction: LSTM can better capture long-term dependencies by processing time series data, and is suitable for predicting monthly or seasonal load changes. By dividing the data into a training set, a validation set, and a test set, and standardizing the data, the consistency of data distribution during the model training process can be guaranteed, and model training problems caused by data bias can be prevented. Training set: used to train the model so that the model can learn the patterns in the data; Validation set: used for parameter adjustment and model selection, to help evaluate the performance of the model under different hyperparameter settings; Test set: used to finally evaluate the generalization ability of the model and ensure that the model performs well on unseen data. Standardization helps to eliminate the scale differences of different features and improve the convergence speed and prediction accuracy of the model. By optimizing the hyperparameters through the grid search method, the hyperparameter combination with the best model performance can be found, thereby improving the prediction accuracy and stability. The grid search method effectively improves the performance of the model by traversing all possible hyperparameter combinations to ensure that no possible optimal configuration is missed. The hyperparameters of the model (such as tree depth, learning rate, etc.) have a significant impact on the prediction results. The grid search method can automatically select the optimal parameters and improve the performance of each sub-model. By evaluating the impact of different time scale features on the prediction results, it is possible to identify which features are most important for load forecasting, thereby providing a basis for subsequent feature selection and model adjustment. These evaluation results help understand the driving factors of load changes and can provide guidance for subsequent optimization models to ensure that more attention is paid to features that contribute greatly to prediction accuracy in the prediction. By evaluating the prediction accuracy of each sub-model on the validation set, the performance of each sub-model in different prediction cycles can be effectively tested, and the pros and cons of the model can be further determined. It can reveal which models have stronger prediction capabilities at a specific time scale, thereby providing data support for weight allocation and model optimization. By combining model evaluation data and feature weight data, and using an adaptive weight allocation method, the weights of each sub-model can be flexibly adjusted according to the prediction accuracy of each sub-model at different time scales. This method can ensure that sub-models that perform better at a certain time scale receive higher weights, thereby improving the accuracy of the entire load forecasting model, especially when the performance of the model at different time scales is unbalanced, the weights can be adjusted dynamically.By optimizing the model structure according to the determined weight data, the prediction accuracy of the model can be further improved, ensuring that the output of each sub-model can be optimally combined according to the needs of different prediction periods. The model after structural optimization is more streamlined and efficient, which can better reflect the characteristics of load forecasting at different time scales and enhance the generalization and robustness of the model. These steps ensure that the model has high accuracy in short-term, medium-term and long-term forecasts through the construction, optimization and evaluation of multi-level load forecasting models. At the same time, through adaptive weight adjustment and model optimization, not only the performance of the model is improved, but also the model can respond more flexibly to changes in different load forecasting periods. These steps help to generate an accurate, efficient and highly adaptable load forecasting model, providing strong data support for the dispatch and management of power systems.

[0093] Preferably, step S33 includes the following steps:

[0094] The grid search method is used to optimize the hyperparameters of each prediction sub-model in the multi-level load forecasting model according to the model training data, so as to obtain the optimized prediction model. The hyperparameter optimization range of random forest includes the number of trees of 100-500, the maximum depth of 5-30, and the minimum number of leaf node samples of 1-10. The hyperparameter optimization range of XGBoost includes the learning rate of 0.01-0.3, the maximum depth of 3-10, the minimum sum of sub-node weights of 1-10, and the subsampling ratio of 0.6-1.0. The hyperparameter optimization range of LSTM includes the number of hidden layers of 1-3, the number of neurons in each layer of 32-256, the drop rate of 0.1-0.5, and the batch size of 16-128.

[0095] In order to optimize the hyperparameters of each prediction sub-model in the multi-level load forecasting model, the embodiment of the present invention selects the grid search method to gradually optimize the parameters of the random forest, XGBoost and LSTM sub-models. First, in the random forest model, the number of trees is set to an integer range of 100-500, the maximum depth is set to an integer range of 5-30, and the minimum number of leaf node samples is set to an integer range of 1-10. The specific method is: by traversing all possible combinations of the number of trees, the maximum depth and the minimum number of leaf node samples, each set of parameters is trained and verified, the mean square error (MSE) is calculated on the verification set and the corresponding results are recorded, and finally the parameter combination with the smallest mean square error is selected as the optimal parameter of the random forest model. In the parameter optimization of the XGBoost model, the learning rate is set to a floating point range between 0.01-0.3, the maximum depth range is an integer value of 3-10, the minimum sub-node weight sum is an integer value of 1-10, and the sub-sampling ratio is a floating point range of 0.6-1.0. The specific optimization steps are: train the model for each set of parameter combinations, record the mean square error of the validation set, and select the optimal parameter group by comparing the mean square error under each combination. This optimization process can improve the generalization ability of the XGBoost model in medium-term forecasting and adapt to the complex changes in daily load. In the LSTM model, the number of hidden layers is set to an integer range of 1-3, the number of neurons in each layer is in the range of 32-256, the drop rate is between 0.1-0.5, and the batch size is set to an integer range of 16-128. Through the grid search method, each parameter combination is traversed one by one, the model is trained on the training set for each set of parameters, and the prediction effect is evaluated on the validation set. The parameter combination corresponding to the optimal mean square error is selected, and the optimal hyperparameter combination of LSTM is finally determined. This optimization step enables the LSTM model to adapt to the long-term dependence in weekly scale forecasting. During the entire grid search process, the specific parameter range of each sub-model is searched separately, and the risk of overfitting of the model on the validation set is reduced through multiple cross-validations, and finally the optimal multi-level load forecasting model is obtained.

[0096] The optimization range of the random forest hyperparameters in the present invention is as follows: The number of trees (100-500): By adjusting the number of trees, the complexity of the model can be controlled in the random forest model. Increasing the number of trees can improve the stability and accuracy of the model and reduce the risk of overfitting and underfitting. Through grid search, the optimal number of trees can be found to balance the training time and prediction accuracy. Maximum depth (5-30): The maximum depth controls the complexity of the decision tree. Deeper trees may capture more training data features, but are prone to overfitting. By adjusting the maximum depth, overfitting or underfitting can be effectively avoided, ensuring that the complexity of the tree matches the needs of the actual data. Minimum number of leaf node samples (1-10): Controlling the minimum number of samples of leaf nodes helps to avoid overfitting of the model to noise during training. The most appropriate minimum number of samples can be found through grid search to ensure that each tree can avoid overfitting to a small number of samples when it is constructed. By optimizing these hyperparameters, random forests can provide more accurate and stable prediction results when dealing with high-dimensional data and complex load forecasting problems. After optimization, the model is more robust in different load scenarios, reduces unnecessary complexity, and improves generalization ability. The optimization range of XGBoost hyperparameters: Learning rate (0.01-0.3): The learning rate controls the step size of the model update at each iteration. A lower learning rate helps to refine the training process and reduce the risk of overfitting, but may require more training rounds. Through grid search, the best learning rate can be found to balance training speed and prediction accuracy. Maximum depth (3-10): The depth of the XGBoost tree determines the complexity of the tree. The greater the depth, the more complex the model is and can capture more nonlinear relationships. Optimizing the maximum depth can help the model adapt to the complex patterns that may exist in load forecasting while avoiding overfitting. Minimum sum of child node weights (1-10): Controls the minimum sum of weights required for splitting. Optimizing this parameter can prevent the model from overfitting when dealing with unbalanced data, especially when dealing with noisy data. Subsampling ratio (0.6-1.0): The subsampling ratio determines the proportion of training data for each tree. A lower ratio can increase the diversity of the model and prevent overfitting. By optimizing this parameter through grid search, the best subsampling ratio suitable for load forecasting can be found. By optimizing the hyperparameters of XGBoost, the model can be made more adaptable and generalizable when dealing with power load forecasting problems. XGBoost is particularly effective in dealing with complex nonlinear relationships and interactions between features. The optimized model can more accurately capture the complex patterns of load data. Optimization range of LSTM hyperparameters: Number of hidden layers (1-3): The number of hidden layers of LSTM affects the capacity and complexity of the model. Increasing the number of hidden layers can make the network more capable of capturing deeper time series features, but it may also increase the difficulty of training. Grid search can be used to determine the optimal number of hidden layers to ensure that the model can effectively capture long-term dependencies.Number of neurons per layer (32-256): The number of neurons per layer directly affects the complexity and fitting ability of the model. By optimizing this parameter, it can be ensured that the LSTM model has sufficient representation ability to capture the long-term temporal dependency of load data while avoiding overfitting during training. Dropout rate (0.1-0.5): The dropout rate is used to reduce overfitting of the neural network. An appropriate dropout rate can improve the generalization ability of the model. A grid search can be used to find a suitable dropout rate to make model training more robust. Batch size (16-128): The batch size determines the amount of data passed to the network during each training. Smaller batches can provide more updates but may cause unstable training; larger batches can speed up training but may cause overfitting. Optimizing the batch size helps improve training efficiency and control the generalization ability of the model. By optimizing these hyperparameters, LSTM can better handle the temporal dependencies in load data. LSTM is particularly suitable for long-term prediction because it can effectively remember historical information in the time series. The optimized LSTM model can perform more accurately and stably when handling long-term dependencies in power load forecasting. By optimizing these hyperparameters through the grid search method, each sub-model (random forest, XGBoost, LSTM) can maximize its advantages in the tasks it excels in, thereby improving the prediction accuracy and stability of the entire multi-level load forecasting model. The load change characteristics at different time scales will be captured more accurately, and the model will not only be more agile in short-term predictions, but also maintain a high prediction ability in medium-term and long-term predictions. The optimized model has stronger generalization ability and robustness, and can provide reliable support in different load forecasting scenarios.

[0097] Preferably, step S4 comprises the following steps:

[0098] Step S41: obtaining current operation data of the thermal power plant;

[0099] The embodiment of the present invention obtains the current operating data of the thermal power plant, and needs to collect key real-time parameters from multiple monitoring systems. These data include the boiler steam pressure and temperature, turbine power output and real-time operating parameters of each generator set read in real time from the unit operation monitoring system; external environmental data such as ambient temperature are obtained from the meteorological station; and the current grid frequency and demand change information are extracted from the power grid dispatching system. The above parameters are integrated through the real-time data interface, and the data is synchronized to the database to ensure that each operating data is complete and meets the time requirements, forming the current operating data that can be used for prediction.

[0100] Step S42: using the load forecasting model to call the corresponding forecasting sub-model based on different forecasting period requirements for the current operation data, and integrating the forecasting data through the model fusion weight to obtain the initial forecasting result data;

[0101] The embodiment of the present invention uses a load forecasting model to perform forecasting operations on the current operating data based on different forecasting cycle requirements. Specifically, the current operating data is input into the short-term, medium-term, and long-term forecast sub-models (random forest model, XGBoost model, and LSTM model, respectively) according to time periods such as hours, days, and weeks. After obtaining the prediction results of each sub-model, the model fusion weight is used for weighted integration. The setting of the model fusion weight is jointly determined by the feature importance and model accuracy during the previous training. For example, the short-term prediction results account for 40%, the medium-term prediction accounts for 35%, and the long-term prediction accounts for 25%. The integrated result is the initial prediction result data, which reflects the load change trend of the thermal power plant at different time scales and provides basic data for the detection of subsequent sudden changes in working conditions.

[0102] Step S43: Perform sudden change operating condition detection according to the current operating data to obtain a sudden change operating condition feature vector, wherein the sudden change operating condition feature vector includes the output power change rate of the generator set, the boiler steam parameter change rate, the grid frequency change rate and the ambient temperature change rate.

[0103] The embodiment of the present invention detects sudden operating conditions based on current operating data. The specific operation is to calculate the rate of change of the current operating data relative to the previous time period, extract the rate of change of the output power of the generator set, the rate of change of the boiler steam parameters, the rate of change of the grid frequency and the rate of change of the ambient temperature, and generate a sudden operating condition feature vector. Set the change rate threshold of each feature, for example, the threshold of the change rate of the output power of the generator set is set to 5%, the threshold of the change rate of the boiler steam parameters is set to 3%, the grid frequency change rate is 0.2%, and the ambient temperature change rate is 1%. Once any parameter in the sudden operating condition feature vector exceeds the corresponding threshold, it is marked as a sudden operating condition, indicating the need for adaptive adjustment of the subsequent model weights. This detection step can effectively identify emergencies in load forecasting and improve the adaptability of the prediction model in practical applications.

[0104] The present invention obtains the current operation data of the thermal power plant, which is the basis for accurate load forecasting. Including data such as the power of the generator set, boiler steam parameters, turbine status, grid frequency and environmental factors, these real-time data provide reliable input for subsequent prediction models. By collecting real-time operation data, the current working state of the power plant can be accurately reflected, ensuring that the load forecasting model can be dynamically adjusted in combination with actual operating conditions, thereby improving the real-time and accuracy of the prediction results. By calling prediction sub-models (such as random forests, XGBoost, LSTM, etc.) based on different time scales (short-term, medium-term, and long-term), it is possible to flexibly select a suitable prediction model according to the actual load forecasting needs of the power plant. This multi-model collaborative working mode makes the prediction results more accurate in different cycles; random forests quickly process recent load fluctuations and provide higher short-term prediction accuracy; XGBoost can combine important features in historical data to provide robust medium-term predictions; LSTM models provide accurate support for long-term load forecasting by capturing long-term dependencies in time series. By merging the prediction results of multiple sub-models according to weights, the advantages of each model can be effectively integrated, the possible deviations of a single model can be reduced, and more accurate and stable initial prediction results can be obtained. The performance differences of different models at different time scales are balanced through fusion to ensure the stability of the comprehensive results. By weighted fusion, the impact of overfitting or underfitting of a single model can be reduced, further improving the overall prediction performance. By detecting sudden changes in operating conditions, abnormal changes under current operating conditions can be identified in a timely manner. For example, a sudden change in the output power of a generator set may indicate system load fluctuations or equipment failures, while abnormal changes in boiler steam parameters, grid frequency, ambient temperature, etc. may indicate unstable operation or external interference factors in the system. Sudden change condition detection can provide early warning signals for power plants, detect potential abnormalities or failure risks in advance, and help operators make timely adjustments or take necessary measures to avoid downtime or efficiency loss caused by emergencies. The sudden change condition feature vector provides necessary support for subsequent data processing and decision-making. It contains the rate of change of different operating parameters, which can help the model identify which factors have a greater impact on load forecasting at a specific moment, thereby optimizing the input of the prediction model and improving the accuracy and response speed of the prediction. Through these three steps, the entire load forecasting system can detect sudden changes in operating conditions in a timely manner based on the real-time operating data of the thermal power plant, combined with the forecasting capabilities of multiple models, thereby improving the accuracy of the forecast and the flexibility of the response. Through the detection of sudden operating conditions, potential risks can be identified in advance to prevent possible abnormalities from affecting the forecast results, ensuring that load forecasts can provide accurate early warnings and decision support under various operating conditions. These steps help improve the operational stability of power plants, reduce downtime, and improve the reliability of load forecasts.

[0105] Preferably, step S43 includes the following steps:

[0106] Step S431: Calculate the time series change rate of the output power of the generator set according to the current operation data, so as to obtain power change characteristic data, wherein the time series change rate includes the instantaneous change rate, the short-term change trend and the power fluctuation amplitude;

[0107] The embodiment of the present invention calculates the rate of change of the time series of the output power of the generator set according to the current operating data. The specific operation is to first calculate the instantaneous rate of change of adjacent sampling points by continuously collecting the output power data of the generator set. For example, when the sampling time interval is 1 minute, the power change rate per minute is calculated; secondly, the short-term change trend is calculated through a sliding time window. The commonly used time windows are 5 minutes and 15 minutes, and the trend of short-term power growth or decline is obtained; finally, the power fluctuation amplitude is calculated based on the maximum and minimum values ​​of the sampled data. The output of this step is power change characteristic data, which can characterize the dynamic change characteristics of the power generation load and is suitable for detecting sudden power fluctuations.

[0108] Step S432: According to the current operation data, the dynamic change of the boiler steam parameters based on the change rate of the main steam pressure and temperature and the stability index are performed to obtain the steam parameter characteristic data;

[0109] The embodiment of the present invention analyzes the dynamic changes of boiler steam parameters based on current operating data. The specific method is to extract real-time data of main steam pressure and temperature, first calculate the rate of change, for example, in units of temperature change (°C / min) and pressure change (MPa / min) per minute; at the same time, calculate the stability index of the steam parameter based on the set stability threshold, and if the parameter changes greatly in a short period of time, it indicates an unstable state. The rate of change and stability index in each time period are combined to obtain steam parameter characteristic data, which is used to reflect the rapid adjustment of boiler load and help detect sudden changes in system operating conditions.

[0110] Step S433: Analyze the real-time fluctuation of the power grid frequency according to the current operation data, calculate the frequency deviation rate and the change rate, and thus obtain the frequency characteristic data;

[0111] The embodiment of the present invention performs real-time fluctuation analysis on the power grid frequency in the current operation data. First, the second-level data of the power grid frequency is obtained, and the frequency deviation rate per second is calculated, that is, the difference ratio with the standard frequency (for example, 50Hz); secondly, the frequency change rate is calculated, which indicates the degree of frequency fluctuation within one second, in units of Hz / s. The frequency characteristic data obtained by the above calculation can be used to determine whether the power grid has a sudden frequency change. If the frequency deviation rate is high or the change rate is large, it means that the power grid has a load shock or other abnormal conditions.

[0112] Step S434: obtaining dynamic change information of the ambient temperature, and performing temperature change rate, temperature gradient and fluctuation characteristic analysis to obtain temperature characteristic data;

[0113] The embodiment of the present invention performs multi-dimensional analysis on temperature data by acquiring dynamic change information of ambient temperature. First, the rate of change of ambient temperature, that is, the temperature change per hour (°C / h), is calculated; second, the temperature gradient is calculated, such as the temperature difference at different locations around the thermal power plant; finally, the temperature fluctuation characteristics are analyzed, such as the temperature difference fluctuation between day and night. The temperature characteristic data obtained through the above analysis can reflect the impact of environmental factors on the operation of the power plant, especially in extreme weather conditions.

[0114] Step S435: standardize the power variation characteristic data, steam parameter characteristic data, frequency characteristic data and temperature characteristic data, and construct a multi-dimensional characteristic space to obtain a sudden change operating condition characteristic vector.

[0115] The embodiment of the present invention performs standardization processing on power change characteristic data, steam parameter characteristic data, frequency characteristic data and temperature characteristic data. The specific method is to perform mean normalization or Z-score standardization on each characteristic data to ensure that each data is in the same scale and eliminate unit differences. Then, the standardized characteristic data is integrated into a multi-dimensional feature vector space, and a mutation condition characteristic vector is constructed to provide a multi-dimensional and uniform scale feature expression for mutation detection, thereby improving the accuracy of condition mutation identification.

[0116] The present invention can deeply analyze the specific characteristics of power fluctuations by calculating the time series change rate of the output power of the generator set. The instantaneous change rate helps to capture rapidly changing fluctuations, the short-term change trend reveals the short-term pattern of power changes, and the power fluctuation amplitude reflects the intensity of power fluctuations. These characteristics help to timely identify sudden load changes or equipment failures. Through the time series change rate, it is possible to identify abnormal conditions of power changes in the generator set, such as abnormal power fluctuations or short-term trends that do not conform to the normal working mode. It is particularly important for the detection of sudden operating conditions, which can help to diagnose potential operating problems at an early stage. Through accurate power change characteristic data, it can provide powerful input for the model, support further operating condition change analysis, and optimize the load prediction model and mutation detection model. The change rate and stability index of boiler steam pressure and temperature are key factors in evaluating the operating status of the boiler. By dynamically monitoring the changes in these parameters, abnormal fluctuations in the boiler system can be discovered in time to prevent potential steam system failures, such as excessive pressure or abnormal temperature rise. Analyzing the change rate and stability of steam parameters can help identify the transition or unstable stage of the boiler system, which is crucial for improving power generation efficiency, extending equipment life, and reducing failure rates. The dynamic change data of steam parameters is an important basis for judging sudden operating conditions. By monitoring these changes, the potential risk of boiler failure can be identified early and early warning can be provided. Grid frequency is an important indicator of power system stability. By analyzing the frequency fluctuation, it is possible to timely identify whether the grid is in normal operation. The frequency deviation rate and change rate provide quantitative characteristics of grid frequency changes, helping to determine whether the grid load fluctuation exceeds the normal range. Frequency fluctuation may be a manifestation of unbalanced load in the power system. By monitoring frequency changes, potential load fluctuations or unstable power supply problems can be discovered, thereby avoiding instability or shutdown of power equipment caused by grid frequency fluctuations. Frequency characteristic data can provide important clues for sudden condition detection. Abnormal changes in frequency deviation and change rate may indicate sudden problems in the grid or unit, such as sudden load increase, failure or equipment status change. Ambient temperature has a significant impact on the operating status of thermal power plants, especially the operating efficiency of boilers and generator sets. Real-time acquisition of ambient temperature and analysis of change rate and gradient can reflect the impact of the external environment on the operation of generator sets and equipment, especially the impact under seasonal changes or sudden climate conditions. Abnormal temperature change rate and gradient may mean the impact caused by external weather or climate change. This information helps to identify equipment performance fluctuations or abnormal operation caused by environmental changes in advance, so as to take preventive measures. Temperature change characteristic data provides external factors to support sudden change condition detection. Sudden changes in ambient temperature may cause abnormal system load and equipment response, so it is an important factor that cannot be ignored in sudden change condition detection.Through standardization, the influence of different feature dimension differences on the analysis is eliminated, so that different types of features (power changes, steam parameters, frequency, temperature) can be compared and integrated on the same scale. This helps to improve the stability and accuracy of the model. The standardized data can form a unified, multi-dimensional feature space, which contains comprehensive information on a variety of key operating parameters. This provides rich data support for subsequent mutation condition detection models and helps to detect more complex and diverse abnormal conditions. The construction of a multi-dimensional feature space enables the model to analyze the current operating status from multiple angles, thereby improving the accuracy of detecting anomalies and predicting mutation conditions. This comprehensive analysis method enhances the system's ability to respond to complex and sudden events.

[0117] Preferably, step S5 comprises the following steps:

[0118] Step S51: comparing the change rate of each parameter in the sudden change operating condition feature vector with its corresponding preset threshold value, so as to obtain sudden change operating condition judgment data;

[0119] The embodiment of the present invention compares the change rate of each parameter in the sudden condition feature vector with its corresponding preset threshold value, and first sets the threshold value of each parameter according to historical data and equipment specifications. For example, the threshold value of the change rate of the output power of the generator set is ±2% / min, the threshold value of the change rate of the boiler steam parameter is ±5% / min, the threshold value of the change rate of the grid frequency is ±0.2 Hz / min, and the threshold value of the change rate of the ambient temperature is ±0.5°C / min. Then, the change rate of each parameter in the sudden condition feature vector is compared with the threshold value set therefor. If the change rate of any parameter exceeds the threshold value, the condition is considered to be a sudden condition. The output of this step is the sudden condition judgment data, indicating whether there is a sudden change in the condition.

[0120] Step S52: when the sudden change condition judgment data shows that the change rates of all parameters do not exceed the preset threshold, the initial prediction result data is directly used as the load prediction data;

[0121] In the embodiment of the present invention, when the sudden change working condition judgment data shows that the change rate of all parameters does not exceed the preset threshold, it means that the working condition is stable and no sudden change occurs. Therefore, the initial prediction result data is directly used as the load prediction data without additional adjustment. In this case, the prediction process maintains the original load prediction model without any intervention, ensuring the consistency and stability of the prediction.

[0122] Step S53: When the sudden operating condition judgment data shows that the change rate of any parameter exceeds the preset threshold, it is determined to be a sudden operating condition, and the weight of the load forecasting model is adaptively adjusted according to the sudden operating condition feature vector, so as to obtain the optimized model fusion weight data;

[0123] In the embodiment of the present invention, when the sudden operating condition judgment data shows that the change rate of any parameter exceeds a preset threshold, it indicates that a sudden operating condition has occurred in the system. At this time, the weight of the load forecasting model is adaptively adjusted through the sudden operating condition characteristic vector. Specifically, according to each parameter in the sudden operating condition characteristic vector, the weight coefficients of different prediction sub-models are adjusted. For example, if the power change rate of the generator set is large, the weight of the short-term forecasting model is increased; if the steam parameter changes drastically, the weight of the medium-term forecasting model is increased. The adjusted weight data optimizes the load forecasting model through weighted fusion to better adapt to the impact of sudden operating conditions, and outputs the optimized model fusion weight data.

[0124] Step S54: Perform weighted fusion processing on the initial prediction result data according to the optimization model fusion weight data to obtain load prediction data.

[0125] The embodiment of the present invention performs weighted fusion processing on the initial prediction result data according to the optimized model fusion weight data. The specific method is to perform weighted averaging on the short-term, medium-term and long-term prediction results according to the weights of each prediction sub-model. For example, the short-term model weight is 0.5, the medium-term model weight is 0.3, and the long-term model weight is 0.2. The final load forecast data is obtained by weighted averaging. This process ensures that when sudden working conditions occur, the prediction results can more accurately reflect the current working condition changes, thereby improving the accuracy and adaptability of load forecasting.

[0126] The present invention can detect in real time whether an abnormal operating condition has occurred by comparing each parameter (such as power change, steam parameter change, frequency fluctuation, etc.) in the sudden change operating condition feature vector with a preset threshold. This comparison mechanism can effectively determine which parameters exceed the normal fluctuation range, thereby timely identifying potential abnormal conditions and ensuring timely response of the system. This step provides a clear judgment basis for the load forecasting model and avoids the possibility of misjudgment and missed judgment. By setting a suitable threshold, normal operating conditions and sudden change operating conditions can be accurately distinguished, thereby better guiding subsequent prediction adjustments. By obtaining sudden change operating condition judgment data, a reliable basis is provided for the next step of model adjustment and weighted fusion, and the prediction method can be dynamically adjusted to improve the accuracy and adaptability of the prediction. If the change rate of all parameters does not exceed the preset threshold, it means that the system is running stably and no abnormal conditions have occurred. In this case, the initial prediction result data is directly used to avoid unnecessary complex adjustment processes. This can save computing resources and improve the response speed and prediction efficiency of the system. When the system is running stably, the initial prediction results are usually optimized, so the direct use of the initial prediction results can ensure the efficiency and accuracy of load prediction, while reducing unnecessary model adjustments and ensuring the smooth operation of the system. This step avoids unnecessary intervention in the model when there is no obvious sudden working condition, thereby ensuring the reliability of the load forecasting system in a stable state. When the change rate of any parameter exceeds the preset threshold, the system can immediately determine it as a sudden working condition and dynamically adjust the load forecasting model. This adaptive adjustment can help the system cope with emergencies, such as equipment failures, load surges, etc., thereby improving the reliability and real-time performance of the forecast results. By adjusting the model weights according to the sudden working condition feature vector, the role and influence of each sub-model in the forecasting process can be dynamically changed, making the model more flexible and adaptable. This process enhances the system's ability to respond to sudden working conditions and improves the robustness of the overall forecasting model. By optimizing the model weights in combination with the sudden working condition feature vector, the forecasting strategy can be adjusted according to the actual working condition changes to ensure that accurate load forecasts can still be made under abnormal working conditions. By weighted fusion of the initial forecast result data and combining the optimized model weights, the contribution of each sub-model can be dynamically adjusted according to the changes in the actual working conditions. This enables the final load forecasting data to combine the advantages of each model and provide more accurate forecasting results. Weighted fusion can adjust the forecast results according to the sudden changes in working conditions in a variety of different situations. When the system is stable, the initial forecast model contributes more; when unexpected conditions occur, the optimized weights will strengthen the influence of specific models to cope with complex and unforeseen situations. Through the optimized fusion weighting method, the model can not only dynamically adjust according to different working conditions, but also improve the performance of the system in load fluctuations. Whether in normal operation or unexpected conditions, the forecast results can accurately reflect the actual load demand, avoiding over-forecasting or under-forecasting.These steps greatly improve the accuracy, flexibility and responsiveness of the load forecasting model through sudden operating condition detection and model adaptive adjustment. Under normal operating conditions, the prediction is efficient and simple; when sudden operating conditions occur, the model is quickly optimized to adapt to new operating conditions. Ultimately, this dynamic adjustment method ensures that the load forecasting system can remain efficient and accurate under different operating conditions, thereby improving the stability and operating efficiency of the power system.

[0127] The present invention also provides a thermal power plant load prediction system, which is used to execute the thermal power plant load prediction method mentioned above. The thermal power plant load prediction system comprises:

[0128] The data acquisition and preprocessing module is used to obtain the real-time operation data of the thermal power plant, including historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data; the real-time operation data is processed for missing data based on the time series interpolation method, and abnormal value detection and correction are performed to obtain operation preprocessing data;

[0129] A time feature extraction module is used to extract multi-scale time features from the running preprocessing data to obtain time feature data, wherein the multi-scale time features include hourly features, daily features, and weekly features;

[0130] The prediction model building module is used to build a multi-level load prediction model based on time characteristic data, wherein the multi-level load prediction model includes a short-term prediction model based on random forest, a medium-term prediction model based on XGBoost, and a long-term prediction model based on long short-term memory network; the multi-level load prediction model is trained using running preprocessed data, and the model weight coefficient is determined according to the characteristics of different prediction periods to obtain the load prediction model;

[0131] The real-time load prediction module is used to use the load prediction model to perform load prediction based on the pre-acquired current operating data to obtain the initial prediction result data; perform sudden change condition detection based on the current operating data to obtain the sudden change condition feature vector, where the sudden change condition feature vector includes the output power change rate of the generator set, the boiler steam parameter change rate, the grid frequency change rate and the ambient temperature change rate;

[0132] The adaptive prediction adjustment module is used to adaptively adjust the weight of the load prediction model when the change rate of any parameter in the sudden change condition characteristic vector exceeds the preset threshold, and weightedly fuse the prediction results of each model according to the adjusted weight coefficient to obtain the load prediction data; when the change rate of the parameters used in the sudden change condition characteristic vector does not exceed the preset threshold, the initial prediction result data is used as the load prediction data.

[0133] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0134] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for predicting load of a thermal power plant, characterized in that: The following steps are involved: Step S1: obtaining real-time operation data of the thermal power plant, including historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data; The real-time operation data is processed for missing data based on the time series interpolation method, and outliers are detected and corrected to obtain operation preprocessing data; Step S2: extracting multi-scale time features from the running preprocessing data to obtain time feature data, wherein the multi-scale time features include hourly features, daily features, and weekly features; Step S3: construct a multi-level load forecasting model based on the time characteristic data, wherein the multi-level load forecasting model includes a short-term forecasting model based on random forest, a medium-term forecasting model based on XGBoost, and a long-term forecasting model based on long short-term memory network; train the multi-level load forecasting model using the running preprocessing data, and determine the model weight coefficient according to the characteristics of different forecast periods to obtain the load forecasting model; Step S4: using the load prediction model to perform load prediction based on the pre-acquired current operating data to obtain initial prediction result data; performing sudden change operating condition detection based on the current operating data to obtain a sudden change operating condition feature vector, wherein the sudden change operating condition feature vector includes the output power change rate of the generator set, the boiler steam parameter change rate, the grid frequency change rate and the ambient temperature change rate; Step S5: When the change rate of any parameter in the sudden change condition characteristic vector exceeds the preset threshold, the weight of the load forecasting model is adaptively adjusted, and the prediction results of each model are weighted and fused according to the adjusted weight coefficient to obtain the load forecasting data; when the change rate of all parameters in the sudden change condition characteristic vector does not exceed the preset threshold, the initial prediction result data is used as the load forecasting data; Step S1 includes the following steps: Step S161: performing data missing detection on the real-time operation data and performing data missing processing to obtain complete operation data, wherein the data missing processing specifically uses linear interpolation to process missing data less than 1 hour, and uses cubic spline interpolation to process missing data greater than 1 hour; Step S162: performing outlier detection based on the Z-score method and the isolation forest algorithm on the complete operation data, and marking the abnormal data points at the identified locations, thereby obtaining outlier marked data, wherein the outlier marked data includes compound marked abnormal data and single marked abnormal data; Step S163: Evaluate the degree of data deviation between the single marked abnormal data and the previous and next time windows, and perform a range check based on the physical constraints of the equipment operation, so as to obtain marked abnormal detection data, wherein the previous and next time windows are specifically ±15 minutes; Step S164: screening the single-label abnormal data for abnormal data points according to the labeled abnormal detection data, thereby obtaining abnormal value labeled screening data; Step S165: merging the composite-marked abnormal data and the abnormal value marked screening data into comprehensive abnormal value data; Step S166: According to the comprehensive abnormal value data, the complete operation data is corrected based on the moving median replacement method, and data smoothing is performed to obtain operation preprocessing data.

2. The method for predicting the load of a thermal power plant according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: extracting historical power generation load curves and data points from the thermal power plant database to obtain historical load data; Step S12: Collecting boiler steam parameters, turbine parameters and overall output power of the generator set in the unit operation monitoring system, thereby obtaining unit operation parameter data; Step S13: Obtain environmental parameter data through a weather station, including ambient temperature, humidity and air pressure data; Step S14: acquiring power grid dispatching data through the power grid dispatching center, including planned power generation data and power grid frequency data; Step S15: merging historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data into real-time operation data; Step S16: Perform data missing processing based on the time series interpolation method on the real-time operation data, and detect and correct abnormal values ​​to obtain operation preprocessing data.

3. The method for predicting the load of a thermal power plant according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: extracting hourly features of the load curve in the hourly time window of the operation preprocessing data, thereby obtaining hourly feature data, wherein the hourly features include hourly average load, load change rate, peak-to-valley difference, and load fluctuation index; Step S22: extracting features from the operation preprocessing data based on the daily cycle, thereby obtaining daily feature data, wherein the daily feature data includes daily maximum load, daily minimum load, daily load rate, daily peak-to-valley difference, and daily load curve shape features; Step S23: Perform weekly scale feature analysis on the operation preprocessing data to extract weekly load variation rules, weekly load distribution characteristics and weekly cycle characteristics, thereby obtaining weekly feature data; Step S24: performing correlation analysis on the hourly feature data, the daily feature data, and the weekly feature data to identify the correlation relationship between features of different time scales, thereby obtaining feature correlation data; Step S25: sorting the importance of each time scale feature according to the feature association data, and performing feature dimension reduction and combination based on the principal component analysis method, so as to obtain time feature data.

4. The method for predicting the load of a thermal power plant according to claim 3, characterized in that: Step S3 includes the following steps: Step S31: construct short-term, medium-term and long-term load forecasting sub-models based on time characteristic data, wherein the short-term forecast adopts the random forest algorithm, the medium-term forecast adopts the XGBoost algorithm, and the long-term forecast adopts the long short-term memory network algorithm, thereby obtaining a multi-level load forecasting model; Step S32: Divide the running preprocessed data into a training set, a validation set, and a test set in a ratio of 8:1:1, and standardize the data to obtain model training data; Step S33: using a grid search method to optimize the hyperparameters of each prediction sub-model in the multi-level load prediction model according to the training set in the model training data, thereby obtaining an optimized prediction model; Step S34: evaluating the importance of the features of each prediction period in the optimized prediction model, calculating the influence of the features of different time scales on the prediction results, and obtaining feature weight data; Step S35: Based on the validation set in the model training data, the prediction accuracy of each sub-model of the optimized prediction model at different prediction time scales is evaluated to obtain model evaluation data; Step S36: Determine the weight coefficients of each sub-model in different prediction periods based on the adaptive weight allocation method according to the model evaluation data and the feature weight data, thereby obtaining model weight data; Step S37: Structural optimization is performed on the optimization prediction model according to the model weight data, thereby obtaining a load prediction model.

5. The method for predicting the load of a thermal power plant according to claim 4, characterized in that: Step S33 includes the following steps: The grid search method is used to optimize the hyperparameters of each prediction sub-model in the multi-level load forecasting model according to the model training data, so as to obtain the optimized prediction model. The hyperparameter optimization range of random forest includes the number of trees of 100-500, the maximum depth of 5-30, and the minimum number of leaf node samples of 1-10. The hyperparameter optimization range of XGBoost includes the learning rate of 0.01-0.3, the maximum depth of 3-10, the minimum sum of sub-node weights of 1-10, and the subsampling ratio of 0.6-1.

0. The hyperparameter optimization range of LSTM includes the number of hidden layers of 1-3, the number of neurons in each layer of 32-256, the drop rate of 0.1-0.5, and the batch size of 16-128.

6. The method for predicting the load of a thermal power plant according to claim 5, characterized in that: Step S4 includes the following steps: Step S41: obtaining current operation data of the thermal power plant; Step S42: using the load forecasting model to call the corresponding forecasting sub-model based on different forecasting period requirements for the current operation data, and integrating the forecasting data through the model fusion weight to obtain the initial forecasting result data; Step S43: Perform sudden change operating condition detection according to the current operating data to obtain a sudden change operating condition feature vector, wherein the sudden change operating condition feature vector includes the output power change rate of the generator set, the boiler steam parameter change rate, the grid frequency change rate and the ambient temperature change rate.

7. The method for predicting the load of a thermal power plant according to claim 6, characterized in that: Step S43 includes the following steps: Step S431: Calculate the time series change rate of the output power of the generator set according to the current operation data, so as to obtain power change characteristic data, wherein the time series change rate includes the instantaneous change rate, the short-term change trend and the power fluctuation amplitude; Step S432: According to the current operation data, the dynamic change of the boiler steam parameters based on the change rate of the main steam pressure and temperature and the stability index are performed to obtain the steam parameter characteristic data; Step S433: Analyze the real-time fluctuation of the power grid frequency according to the current operation data, calculate the frequency deviation rate and the change rate, and thus obtain the frequency characteristic data; Step S434: obtaining dynamic change information of the ambient temperature, and performing temperature change rate, temperature gradient and fluctuation characteristic analysis to obtain temperature characteristic data; Step S435: standardize the power variation characteristic data, steam parameter characteristic data, frequency characteristic data and temperature characteristic data, and construct a multi-dimensional characteristic space to obtain a sudden change operating condition characteristic vector.

8. The method for predicting the load of a thermal power plant according to claim 7, characterized in that: Step S5 includes the following steps: Step S51: comparing the change rate of each parameter in the sudden change operating condition feature vector with its corresponding preset threshold value, so as to obtain sudden change operating condition judgment data; Step S52: when the sudden change condition judgment data shows that the change rates of all parameters do not exceed the preset threshold, the initial prediction result data is directly used as the load prediction data; Step S53: When the sudden operating condition judgment data shows that the change rate of any parameter exceeds the preset threshold, it is determined to be a sudden operating condition, and the weight of the load forecasting model is adaptively adjusted according to the sudden operating condition feature vector, so as to obtain the optimized model fusion weight data; Step S54: Perform weighted fusion processing on the initial prediction result data according to the optimization model fusion weight data to obtain load prediction data.

9. A thermal power plant load prediction system, characterized in that: The method for predicting the load of a thermal power plant according to claim 1, wherein the prediction system for the load of a thermal power plant comprises: The data acquisition and preprocessing module is used to obtain the real-time operation data of the thermal power plant, including historical load data, unit operation parameter data, environmental parameter data and power grid dispatching data; the real-time operation data is processed for missing data based on the time series interpolation method, and abnormal value detection and correction are performed to obtain operation preprocessing data; A time feature extraction module is used to extract multi-scale time features from the running preprocessing data to obtain time feature data, wherein the multi-scale time features include hourly features, daily features, and weekly features; The prediction model building module is used to build a multi-level load prediction model based on time characteristic data, wherein the multi-level load prediction model includes a short-term prediction model based on random forest, a medium-term prediction model based on XGBoost, and a long-term prediction model based on long short-term memory network; the multi-level load prediction model is trained using running preprocessed data, and the model weight coefficient is determined according to the characteristics of different prediction periods to obtain the load prediction model; The real-time load prediction module is used to use the load prediction model to perform load prediction based on the pre-acquired current operating data to obtain the initial prediction result data; perform sudden change condition detection based on the current operating data to obtain the sudden change condition feature vector, where the sudden change condition feature vector includes the output power change rate of the generator set, the boiler steam parameter change rate, the grid frequency change rate and the ambient temperature change rate; The adaptive prediction adjustment module is used to adaptively adjust the weight of the load prediction model when the change rate of any parameter in the sudden change condition characteristic vector exceeds the preset threshold, and weightedly fuse the prediction results of each model according to the adjusted weight coefficient to obtain the load prediction data; when the change rate of all parameters in the sudden change condition characteristic vector does not exceed the preset threshold, the initial prediction result data is used as the load prediction data.

Images