A power grid adaptive dispatching method combined with short-term power source and load forecasting

Abstract

The present invention discloses a power grid adaptive scheduling method combined with short-cycle power source and load prediction, and relates to the technical field of power prediction. At present, the inaccuracy of power source and load prediction caused by short-cycle drastic weather changes; the present invention includes: obtaining weather change information and extracting features according to the short-cycle preset time granularity; constructing similar day features according to the time granularity, weather features and current time, and constructing a similar day database based on the extracted historical data; preprocessing the database and aggregating features to obtain similar day aggregate features; inputting weather features into the power grid prediction model to obtain short-cycle power source and load prediction results; and performing power grid adaptive scheduling according to the balance relationship of the prediction results. This technical solution improves the accuracy of short-cycle new energy prediction and short-term load prediction by combining similar day features and stacking LSTM models.

Description

Technical Field

[0001] The present application relates to the technical field of power forecasting, and in particular to a power grid adaptive scheduling method combined with short-cycle power source and load forecasting. Background Art

[0002] As the proportion of renewable energy in the global power system continues to increase, the challenges faced by short-term power source and load forecasting are also gradually increasing. The volatility and unpredictability of energy sources such as wind and solar energy make grid dispatching complicated, especially in short periods (such as hours or minutes) when meteorological changes have a more obvious impact on power demand and supply. Traditional grid dispatching methods are usually based on long-term forecasts and lack the ability to respond to weather changes and power source and load fluctuations in short periods. This leads to insufficient flexibility of the grid in responding to emergencies and power imbalances, which in turn affects the operational stability and power supply reliability of the grid. Existing dispatching methods usually fail to fully utilize the advantages of short-term power source and load forecasting, and fail to effectively respond to the impact of short-term drastic weather changes on power demand and supply. Therefore, how to improve the accuracy of short-term power source and load forecasting and realize grid adaptive dispatching has become a technical problem that needs to be solved in the current power dispatching field. Summary of the invention

[0003] The present application provides a power grid adaptive scheduling method combined with short-cycle power source and load forecasting, aiming to solve the technical problem of inaccuracy in power source and load forecasting caused by short-cycle drastic weather changes.

[0004] In view of the above problems, the present application provides a power grid adaptive scheduling method combined with short-cycle power source and load prediction.

[0005] The present application provides a power grid adaptive scheduling method combined with short-cycle power source and load prediction, the method comprising: obtaining weather change information according to a short-cycle preset time granularity, performing feature extraction on the weather change information, and obtaining weather change features, wherein the short-cycle preset time granularity includes multiple levels of time granularity; constructing similar day features according to the short-cycle preset time granularity, weather change features and current time positioning, extracting historical data based on similar day features to construct a similar day database; preprocessing the similar day database, and performing feature aggregation on the similar day database to obtain similar day aggregation features, wherein the similar day aggregation features have a time granularity identifier; inputting the weather change features into a power grid prediction model to obtain a short-cycle power source and load prediction result, wherein the power grid prediction model is a stacked LSTM model, which is obtained by learning a training data set constructed by similar day aggregation features; and performing power grid adaptive scheduling according to the balance relationship of the short-cycle power source and load prediction results.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] The above-mentioned power grid adaptive scheduling method combined with short-cycle power source and load prediction obtains weather change information according to the multi-level time granularity preset in the short cycle, and extracts features to obtain weather change features; then, in combination with the current time positioning, similar day features are constructed, and a similar day database is constructed by extracting date features similar to similar day features from historical data. This database organizes and aggregates historical data, which not only reduces data redundancy, but also enhances correlation through effective feature aggregation, ensuring that the prediction model can capture the complex nonlinear relationship between weather and power load; then, the extracted weather change features are input into the power grid prediction model constructed using the stacked LSTM model to capture the dynamic fluctuations of power source and load with time and weather changes, and obtain the short-cycle power source and load prediction results; then, based on the short-cycle power source and load prediction results, the power grid can perform adaptive scheduling according to the balance relationship between source and load. This scheduling can not only respond to short-cycle power fluctuations in real time, but also effectively optimize the operation status of the power grid, avoid power imbalance caused by short-cycle drastic weather changes, thereby improving the stability of the power grid and power supply reliability.

[0008] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0010] Figure 1 It is a schematic diagram of the process of the present invention.

[0011] Figure 2 The figure is a schematic diagram of the process of obtaining similar daily aggregation characteristics according to the present invention. DETAILED DESCRIPTION

[0012] The embodiment of the present application solves the technical problem of inaccuracy in power source and load prediction caused by short-term drastic weather changes by providing a power grid adaptive scheduling method combined with short-term power source and load prediction.

[0013] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0014] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules that are not explicitly listed or inherent to these processes, methods, products or devices.

[0015] Examples, such as Figure 1 As shown, the present application provides a power grid adaptive scheduling method combined with short-cycle power source and load prediction, the method comprising:

[0016] According to the short-period preset time granularity, weather change information is obtained, and feature extraction is performed on the weather change information to obtain weather change features, wherein the short-period preset time granularity includes multiple levels of time granularity.

[0017] In an embodiment of the present application, the short-cycle preset time granularity includes multiple time levels such as hours, days, weeks, and months, and each level corresponds to a time granularity, namely, hourly granularity, daily granularity, weekly granularity, monthly granularity, etc.; through these time granularities, weather change information at different time granularities is obtained through meteorological data acquisition equipment, including but not limited to data such as temperature, humidity, wind speed, wind direction, and radiation; for the weather change information at each time granularity, a feature extraction algorithm (such as Fourier transform, time series analysis, etc.) is used to extract weather features that are helpful for power source and load prediction. For example, weather features such as temperature changes, wind speed fluctuations, and air pressure changes in the past few hours are extracted based on the hourly granularity, and weather features such as the average temperature, daytime and nighttime temperature difference, and wind speed trend in the past few days are extracted based on the daily granularity. Through this process, weather patterns at different time granularities can be captured, thereby obtaining weather change features, ensuring that effective feature information can be provided in different time periods, and providing accurate data support for subsequent power grid adaptive scheduling.

[0018] According to the short-period preset time granularity, weather change characteristics and current time positioning, similar day features are constructed, and historical data are extracted based on the similar day features to construct a similar day database.

[0019] In one embodiment, the corresponding time position is determined in the historical database of the meteorological data collection device according to the current time location, in order to ensure that subsequent operations can be performed based on accurate historical data; then, a time period is calculated forward according to a short-cycle preset time granularity (such as hour, day, week, month, etc.) to obtain a past time point. For example, if the current time granularity is hourly granularity, one hour granularity is calculated forward from the current time point; if the time granularity is daily granularity, one day granularity is calculated forward; then, starting from the calculated past time point, historical data corresponding to the current time granularity (such as the past hour or past month) is obtained forward. The meteorological data of the past day are collected, and then the similarity between these historical data and the current weather change characteristics is calculated. The similarity is calculated by measuring the absolute difference between each feature (such as temperature, humidity, wind speed, etc.) to evaluate the degree of deviation. If the deviation between the historical data in a certain time period and the current weather change characteristics is within the preset tolerance range, it is considered that the weather characteristics of this time period are similar to the current weather changes. At this time, the historical data that meets the conditions are added to the similar day database. Through this process, data similar to the current meteorological conditions can be effectively extracted from the historical weather data, providing high-quality input data for subsequent power source and load forecasts.

[0020] The similar day database is preprocessed, and feature aggregation is performed on the similar day database to obtain similar day aggregate features, where the similar day aggregate features have a time granularity identifier.

[0021] In one embodiment, after constructing a similar day database, the data in the similar day database will be preprocessed. The purpose of preprocessing is to clean the data, remove redundant information, outliers or missing values, and ensure the quality and consistency of the data. After the data preprocessing is completed, feature aggregation will be performed on the multidimensional data in the similar day database. Feature aggregation refers to merging weather features from different time granularities (such as hours, days, weeks, and months) to obtain a more comprehensive and streamlined feature set. During the aggregation process, horizontal feature rolling aggregation and vertical feature aggregation can be performed according to the rolling prediction technology, so as to aggregate features of the same granularity to form similar day aggregate features. During the feature aggregation process, each aggregated feature needs to be The corresponding time granularity should be marked. The time granularity identification is to ensure that the characteristic data at different time scales can be distinguished in subsequent analysis. For example, after the temperature change characteristics of a certain hour are aggregated, they will be marked as hourly granularity; the average wind speed characteristics of a day will be marked as daily granularity. The purpose of this is to help the model accurately understand the time background of each aggregated feature and ensure that the role of time scale in prediction and scheduling decisions is fully reflected. Through this step, aggregated features with time granularity identification can be effectively extracted from the similar day database. These aggregated features provide concise and high-quality data input for the subsequent power source and load prediction model, helping the model to better capture the relationship between power demand and weather changes.

[0022] Further, if Figure 2 As shown, the present application provides preprocessing of the similar day database, and feature aggregation of the similar day database to obtain similar day aggregate features, including:

[0023] The similar day database is cleaned and filtered, and a time window is constructed according to a preset time granularity to perform multi-level granular data extraction on the similar day database, and a multi-level data architecture is constructed, in which each level corresponds to a granular time window; based on the multi-level data architecture, the rolling prediction technology is used in the same level to identify the feature relationship of the granular time window and perform horizontal feature rolling aggregation; based on the multi-level data architecture, the horizontal rolling aggregation features of the hierarchical granularity window are vertically aggregated according to the hierarchical neighborhood relationship to obtain the similar day aggregation features.

[0024] Preferably, before performing feature aggregation on the similar day database, the similar day database is first cleaned and filtered. The cleaning process includes removing duplicate data, filling missing values, and processing outliers to ensure the quality of the data, which provides a clean and reliable data basis for subsequent data extraction and aggregation. The purpose of data filtering is to ensure that only data that meets the standards enters the subsequent processing flow; wherein, the removal of duplicate data can be performed using a hash algorithm or a deduplication method based on a unique identifier (time point), the filling of missing values ​​can be performed using a mean interpolation or interpolation method, and the processing of outliers can be performed using an IQR method (quartile method) or a Z-Score method; subsequently, according to the set short-term preset time granularity (such as hours, days, weeks, Month), divide the historical similar daily data into multiple time windows, each time window represents a time period of a granularity. For example, each time window of hourly granularity represents one hour of data, such as one hour of content temperature, humidity, wind speed, etc., each time window of daily granularity represents one day of data, each time window of weekly granularity represents one week of data, and each time window of monthly granularity represents one month of data; the data in each time window will be extracted to form a multi-level data architecture, and the data at each level will contain the feature data under the time granularity of that level for subsequent feature aggregation and analysis; after the multi-level data architecture is built, the rolling prediction technology will be used to identify feature relationships at the same level, which means that According to the data in the time window, the relationship and trend between features are identified, and horizontal feature aggregation is performed; among them, horizontal aggregation refers to the merging of features of adjacent time windows at the same time granularity level. For example, for hourly granularity data, the trend of changes in temperature, humidity, etc. within several consecutive hours will be identified, and these features will be aggregated to reflect more comprehensive weather change characteristics, thereby obtaining the aggregated features of this level and adding them to the horizontal rolling aggregation features; after completing the horizontal aggregation, vertical feature aggregation is performed based on the hierarchical neighborhood relationship in the multi-level data architecture. The purpose of vertical aggregation is to further improve the overall expressiveness of the data by combining the aggregated features of different granularity levels. Vertical aggregation includes data at different levels. For example, the hourly granularity features are mapped to the daily granularity cycles, and then the data are aggregated according to the upstream and downstream connections of time by aggregating the average values ​​of adjacent features. This aggregation can help better capture the feature associations across time scales, thereby improving the model's ability to recognize weather patterns. After horizontal and vertical aggregation, the aggregate features representing similar days are finally obtained. These aggregate features not only cover data of multiple time granularities, but also reflect the multi-level correlation of weather changes, and have high predictive ability. In addition, each similar day aggregate feature has a clear time granularity identifier, ensuring that in the subsequent power source and load forecast, the model can process and analyze according to the features of different granularities.

[0025] Furthermore, the present application provides a method for identifying feature relationships of granular time windows using rolling prediction technology at the same level and performing horizontal feature rolling aggregation, including:

[0026] Based on the granular time window, target feature relationship recognition is performed to determine the relationship overlap rate; based on the granular time window, temporal relationship recognition is performed to determine the time distance; relationship fusion is performed based on the relationship overlap rate and the time distance, and aggregation enhancement weights are configured; features in the granular time window at the same level are weightedly aggregated based on the aggregation enhancement weights to obtain horizontally scrolling aggregation features.

[0027] Optionally, in the same level, target feature relationship recognition is performed based on the set granular time window. The features in each time window (such as temperature, wind speed, humidity, etc.) will be compared with the features in other time windows to identify the correlation between them. The core of feature relationship recognition is to determine the similarity of features in different time windows, and then calculate their relationship overlap rate. This relationship overlap rate refers to the degree of feature similarity between two time windows. The more similar the features are, the higher the overlap rate is. In the process of calculating the relationship overlap rate, the adjacent granular time windows in the level will be combined, and the relationship overlap rate calculation formula will be used. Calculate; where R is the relationship overlap rate of the current granular time window combination, between 0 and 1. The closer it is to 1, the more similar the features are. n is the number of features in the granular time window. is the i-th feature of the first granular time window in the current granular time window combination, is the i-th feature of the second granular time window in the current granular time window combination; after calculating the relationship overlap rate, the temporal relationship will be analyzed, that is, the time distance of each granular time window combination. The smaller the time distance, the closer the data of the two time points are, and the greater the impact may be. Therefore, the time difference between the two time windows will be used as the time distance; after calculating the relationship overlap rate and the time distance, the maximum and minimum value method will be used to normalize the time distance, and then the normalized time distance and the relationship overlap rate will be fused through mean calculation to configure the aggregation enhancement weight of each granular time window combination. The higher the relationship overlap rate and the closer the time window is, the higher the weight; then, The granular time window combinations whose aggregation enhancement weights are greater than or equal to the preset aggregation enhancement weights are weighted aggregated, that is, the features in the two granular time windows are fused together through the aggregation enhancement weights to form an initial horizontal rolling aggregation feature; the above process is repeated until there is no granular time window combination greater than or equal to the preset aggregation enhancement weight, thereby obtaining the horizontal rolling aggregation feature; through horizontal rolling aggregation, more accurate feature data can be obtained to help improve the accuracy of short-term power source and load forecasts. This process not only takes into account the similarity of weather characteristics, but also strengthens the contribution of adjacent time windows to the forecast results through weight configuration, making the model more sensitive and efficient in dealing with short-term fluctuations.

[0028] Furthermore, the present application provides a method for performing vertical feature aggregation on the horizontal rolling aggregation features of the hierarchical granularity window based on the multi-level data architecture according to the hierarchical neighborhood relationship, including:

[0029] Identify the timing alignment relationship of the hierarchical granularity windows; perform hierarchical granularity window conversion mapping based on the timing alignment relationship, and establish the timing alignment mapping relationship of each level of the data architecture; according to the timing alignment mapping relationship, aggregate the average values ​​of adjacent features from small to large according to the hierarchical granularity windows to obtain similar daily aggregate features.

[0030] Optionally, when performing longitudinal feature aggregation, first, identify the temporal alignment relationship between different time granularities to ensure that the time windows between different granularities can be effectively connected and matched when performing multi-level feature aggregation; in this process, the features in all levels are matched one by one according to the timestamp. For example, for hourly and daily granularities, it is identified which time period each hour belongs to, and it is ensured that the hourly data can be mapped to the corresponding daily data period. For daily and weekly granularities, the daily data of every seven days is mapped to the time period of a week to ensure that the data of each week can be aggregated through the temporal alignment relationship; when the temporal alignment relationship between the levels is identified, the time window conversion is performed to map the window data of the lower granularity (such as hours) to the window data of the higher granularity (such as days and weeks) to ensure that the data of different periods can be correctly mapped to the corresponding granularity level; for example, for the conversion from hours to days, the 24-hour data of each day is integrated into a single window of daily granularity, and for the conversion from days to weeks, the daily data of seven consecutive days is aggregated into a window of weekly granularity. ; After completing the hierarchical granularity window conversion, the granularity window data of each level in the multi-level data architecture is mapped to the upper level according to the identified timing alignment relationship, so as to establish the timing alignment mapping relationship of each level of the data architecture to ensure that the time window data of each level can be correctly connected and docked; for example, the data of every 24 hours of the first level is mapped to the day of the second level, and the 7-day data of the second level is mapped to the week of the third level; according to the established timing alignment mapping relationship, the features in each level will be aggregated. For example, when aggregating from hours to days, the aggregated features of the day are obtained by averaging the 24-hour data of a day; when aggregating, it will be ensured that the order is from small to large, that is, the aggregation from hour granularity to day granularity is performed first, and then the aggregation from day granularity to week granularity is performed, and so on. The data of each time window is aggregated according to its timing alignment mapping relationship. The average value of adjacent features is aggregated, and finally a comprehensive similar day aggregation feature reflecting weather change characteristics is formed, which provides accurate historical weather change information for the prediction of power source and load.

[0031] The weather change characteristics are input into a power grid prediction model to obtain a short-term power source and load prediction result, wherein the power grid prediction model is a stacked LSTM model, which is learned through a training data set constructed by similar daily aggregation features.

[0032] In one embodiment, the obtained weather change characteristics are taken as input and passed to the power grid prediction model trained by learning the training data set constructed based on similar day aggregation characteristics for analysis. The power grid prediction model will generate short-term power source and load prediction results based on the learned mapping relationship. These prediction results represent the power demand in the future (such as the next few hours or days), which helps the power grid to dynamically dispatch and allocate resources. Among them, the power grid prediction model adopts a stacked LSTM (Long Short-Term Memory) model, which is a deep learning model suitable for time series data. The stacked LSTM model effectively captures the long-term and short-term dependencies between weather changes and power sources and loads by stacking multiple LSTM layers, thereby learning how weather changes affect future power demand and making predictions on multiple time scales, providing reliable data support for the adaptive dispatch of the power grid.

[0033] Further, the present application provides inputting the weather change characteristics into a power grid prediction model, which also includes:

[0034] Construct a multi-layer stacked LSTM model architecture, including an input layer, multiple LSTM layers, a fully connected layer, and an output layer; based on the multi-level time granularity, construct a multi-level training data set and a verification data set according to the similar daily aggregation features; perform multiple LSTM layer training and verification convergence on the multi-level training data set and verification data set respectively to obtain the power grid prediction model.

[0035] Preferably, when constructing a power grid prediction model, first, a multi-layer stacked LSTM model is used to structure the power grid prediction model, including an input layer, multiple LSTM layers, a fully connected layer, and an output layer; wherein the input layer receives weather change feature data that has been feature extracted and aggregated, and these data include temperature, humidity, wind speed, etc., and are organized into time series data according to a set time granularity (such as hours, days, weeks, etc.); multiple LSTM layers are stacked together, and each LSTM layer processes the output of the previous layer and the data at the current time point to capture the short-term dependency between power load and weather changes, and the fully connected layer is used to convert the time series features extracted by the LSTM layer into time series data. The features are mapped to the power load forecast results. The output of the fully connected layer will be the predicted value, that is, the prediction of future power source load. The output layer is used to output the prediction results of the model. In addition, similar daily aggregate features will be divided into training data sets and verification data sets according to multi-level time granularity and prediction targets for training and verification of the model. After completing the architecture of the power grid prediction model and the construction of training data, the training data set is input into the initialized multi-layer stacked LSTM model for forward propagation. The processed weather change feature data is received through the input layer and then passed layer by layer through multiple LSTM layers. Each LSTM layer calculates the hidden state according to the output of the previous layer and the current input. Extract the time series features required for power load prediction; the data passes through the fully connected layer and is passed to the output layer to output the prediction results of the power source and load; after obtaining the prediction results, use the mean square error (MSE) loss function to calculate the loss value between the prediction results and the actual power load, and calculate the gradient of the loss function to the weights of each layer layer by layer through back propagation, and then use the Adam optimizer to adjust the model parameters and optimize the loss function, so that the model can continuously improve the prediction results during the training process. The optimization process will be repeated until the loss function converges or the predetermined number of training times is reached; during the training process, the system terminal uses the verification data set for real-time evaluation to ensure the performance of the model on the verification set There will be no overfitting. When the loss on the validation set tends to be stable and the loss on the training set decreases, the training process stops and the model is considered to have converged. After training, the model is finally evaluated using data not used for training and validation to test its generalization ability. The evaluation results are measured by calculating indicators such as mean square error (MSE) to ensure the accuracy and stability of the model in practical applications. Finally, after training, verification and optimization, a power grid prediction model is generated. Based on the real-time weather change characteristics, the model can accurately predict the power source and load in a short period of time, provide decision support for power grid dispatching, and help power grid operators to predict load and optimize resource allocation.

[0036] Furthermore, the present application provides a method for constructing a multi-level training data set and a verification data set based on the multi-level time granularity and the similar day aggregation features, including:

[0037] Taking the multi-level time granularity as an index, a time granularity identifier matching search is performed in the similar day aggregation features to obtain the similar day aggregation features of each time granularity; according to the prediction target, the similar day aggregation features are subjected to target identification to establish a standardized training data set; based on the standardized training data set, data formatting is performed to convert the data into a format that conforms to the LSTM model input format, including the number of samples, the time step, and the number of features; according to the time series of the similar day aggregation features, the formatted training data set is divided into a training data set and a verification data set, wherein the time series of the training data set is earlier than the time series of the verification data set.

[0038] Optionally, after obtaining the similar day aggregate features, multi-level time granularity is used as an index, and a time granularity identification matching search is performed from the similar day aggregate features, and the aggregate features represented by the same time granularity are integrated together to form similar day aggregate features of each time granularity; then, the similar day aggregate features of each time granularity are identified according to the prediction target. When the prediction target is the power source and load in a certain period of time in the future, the load data of the target time period in the similar day aggregate features of each time granularity will be labeled to establish an initial training data set; in the similar day aggregate features of the hourly granularity, the hourly load average, maximum load, temperature, wind speed, etc. of the past 7 days are used as input features, and the power load in the next 24 hours is used as output features. Output labels, for example, the hourly load average of the past 7 days is 30kW, the maximum load is 40kW, the temperature average is 25°C, the wind speed average is 5m / s, etc.; for similar daily aggregation features with daily granularity, the daily average load, daily maximum load, temperature and wind speed of the past 30 days are used as input features, and the power load of the next day is used as a label. For example, the daily average load of the past 30 days is 35kW, the daily maximum load is 50kW, the daily average temperature is 28°C, and the wind speed average is 8m / s; after determining the initial training data set, the data in the initial training data set is normalized by the maximum and minimum method, and the data is scaled to between 0 and 1 to form a standardized training data set. For example, the input feature temperature in the past 24 hours The range is [20°C, 30°C], and 25°C can be normalized to 0.5; then, the standardized training data set is formatted, that is, the data is converted into a format suitable for LSTM model input. The input of LSTM is usually a three-dimensional array [number of samples, time steps, number of features], where the number of samples is the number of samples in the standardized training data set. For example, from January 1, 2015 to December 31, 2020, there are a total of 365*6=2190 samples (assuming there are 365 days in a year and 6 years), and the time step is the time series length of each sample. For example, the data for the past 7 days can be represented as 7 time steps (each time step represents one day), and the number of features is 1 for each time step. The number of data features in the interval, for example, load, temperature, humidity, etc.; then, the formatted standardized training data set is divided. For time series prediction tasks, the division of training data sets and validation data sets needs to follow the time order, and random division cannot be used. Therefore, data from an earlier time period is used to form the training data set. For example, data from January 1, 2015 to December 31, 2019 is selected, and data from the time period after the training data set is used to form the validation data set. For example, data from January 1, 2020 to June 30, 2020 is selected; through the above steps, high-quality training data can be effectively provided for the power grid prediction model, and the accuracy and generalization ability of the model in real applications can be guaranteed.

[0039] According to the balance relationship of the short-cycle power source and load prediction results, the power grid is adaptively dispatched.

[0040] In one embodiment, after obtaining the short-cycle power source and load prediction results, the power load prediction results and the energy output prediction results in the short-cycle power source and load prediction results are evaluated for the source-load balance relationship, and based on the source-load balance relationship evaluation results, the scheduling path is parsed through a pre-built adaptive scheduling module to determine the nodes that need to be scheduled and the required scheduling amount; subsequently, the power supply strategy is readjusted according to the determined scheduling nodes and the required scheduling amount, and the power grid is adaptively scheduled according to the adjusted power supply strategy to avoid power shortages or surpluses and ensure the balance between power load and power supply, thereby achieving efficient, stable and reliable power grid operation.

[0041] Furthermore, the present application provides a method for performing adaptive grid scheduling according to the balance relationship of the short-cycle power source and load prediction results, including:

[0042] According to the short-cycle power source-load prediction results, the power load prediction results and the energy output prediction results are obtained; the source-load balance relationship is evaluated according to the power load prediction results and the energy output prediction results to obtain the source-load balance relationship; a target power grid topology structure is established, which includes power generation energy nodes and load center nodes; node dependency relationships are identified and extracted according to the target power grid topology structure, and an adaptive scheduling module is constructed; balance compensation is performed according to the source-load balance relationship to obtain a scheduling target; the scheduling target is used as input, and a scheduling path is parsed through the adaptive scheduling module to obtain adaptive scheduling information, and the adaptive scheduling information includes scheduling nodes and scheduling adjustment amounts.

[0043] Preferably, after obtaining the short-cycle power source and load prediction results, the power load and energy output prediction results for the future time period will be extracted from the short-cycle power source and load prediction results. These prediction results are the basis for subsequent source-load balancing, scheduling optimization and power grid control; then, according to the power load prediction and energy output prediction in the prediction results, the source-load balance relationship is calculated. The source-load balance refers to the degree of matching between the power load demand and the energy supply. The difference between the predicted power output in the energy output prediction result and the predicted load demand in the power load prediction result is calculated. If the calculation result is a positive value, that is, the predicted power output is greater than the load demand, it means that the power supply is in excess, and it may be necessary to avoid waste by reducing the power generation or the charging amount of the energy storage device. If the result is a negative value, that is, the predicted power output is less than the load demand, it means that the power demand exceeds the supply capacity, and it may be necessary to increase the power generation or enable the backup power supply, or use the energy storage device to supplement the supply; based on the calculation results, the balance relationship between the source and the load is determined, for example, supply and demand balance, overload, and energy shortage, and the corresponding supply and demand ratio is configured for the source-load balance relationship, that is, the ratio of the predicted power output to the predicted load demand, which is used to measure whether the source and the load are balanced.

[0044] When performing adaptive grid dispatch, a target grid topology is also established. This target grid topology represents the connection relationship between each power generation energy node and the load center node in the power network. Among them, the power generation energy node is the power station or energy input point in the power grid, including traditional thermal power, hydropower, nuclear power plants, and renewable energy wind power, photovoltaic and other power generation facilities. The load center node is the main area or region of power consumption, such as urban centers, industrial parks, etc. The establishment of the target grid topology can help determine the power transmission path, the output distribution of the power generation node and the power demand of the load center, thereby providing structured information for subsequent adaptive dispatch. Subsequently, based on the target grid topology, node dependency is identified and extracted. The node dependency reflects the relationship between each node in the power grid, including the power transmission path from the power generation node to the load center node, and the power transmission capacity limit between nodes. Based on the node dependency, an adaptive dispatch module is constructed through the Laplace matrix. This module will dynamically adjust the power transmission between each node in the power grid according to the source-load balance relationship, node dependency and grid topology to optimize the operation efficiency of the power grid.

[0045] Through the previous calculation of the source-load balance relationship, it is possible to understand whether the power grid is in a balanced state. If there is an imbalance (such as excessive load or excess electricity), balancing compensation is required; at this time, according to the source-load balance relationship, a scheduling target is determined to represent the optimal operating state of the power grid in the future. This target includes how to adjust the power flow between nodes to achieve a balance between power demand and supply; for example, when it is predicted that the load is too large, it may be necessary to increase power generation or supplement power through energy storage equipment; when there is excess electricity, it may be necessary to reduce power generation or adjust the energy storage charging and discharging strategy; after determining the scheduling target, the scheduling target is used as input to use the constructed adaptive scheduling module to perform scheduling path analysis. This process includes determining the starting point from the power generation node The optimal power transmission path and scheduling scheme from the point to the load center node; according to the target power grid topology, node dependencies, and the balance between power demand and supply, the adaptive scheduling module calculates the power flow difference of each path. For example, a power generation node may provide power to different load centers through multiple transmission lines. The scheduling path analysis helps determine the best way to transmit power; the results of the scheduling path analysis will generate adaptive scheduling information, including the scheduling amount of the power generation node (such as increasing or decreasing the power generation output), the charging and discharging amount of the energy storage equipment, and the power scheduling demand of the load center (such as increasing or decreasing the power input). These scheduling information will be fed back to the power grid system in real time to guide the operation of the power grid and ensure the stable and efficient operation of the power supply.

[0046] Furthermore, the present application provides a method for identifying and extracting node dependencies according to the target power grid topology structure and constructing an adaptive scheduling module, including:

[0047] Based on the target power grid topology, the connection relationship between each node is obtained to construct an adjacency matrix; based on the target power grid topology, the number of adjacent nodes of the node is obtained to construct a degree matrix; difference calculation is performed based on the adjacency matrix and the degree matrix to obtain the dependency relationship of each node; based on the dependency relationship of each node, the equilibrium identification relationship between the scheduling path and the scheduling coefficient is fitted to construct the adaptive scheduling module.

[0048] Optionally, based on the target power grid topology, the connection relationship between each node in the power grid (including power generation nodes and load nodes) is determined. The target power grid topology is composed of multiple power generation nodes, load center nodes and power transmission lines between them. The topology is in the form of a graph, in which the nodes represent power generation and load points, and the edges represent power transmission lines between the nodes. Based on this connection relationship, an adjacency matrix A is constructed. In the adjacency matrix, if there is a power transmission path between node i and node j, then A in the adjacency matrix ij =1, otherwise A ij=0; Similarly, based on the target power grid topology, the number of adjacent nodes of each node is also identified to construct the degree matrix D. In the degree matrix, each diagonal element D i represents the degree of node i. For example, if node 1 is connected to node 2 and node 3, then the degree of node 1 is 2. 1 =2; Then, according to the adjacency matrix A and the degree matrix D, the Laplace matrix L is calculated. The Laplace matrix is ​​used to represent the connection relationship between nodes and the current flow path; according to the Laplace matrix, the dependency difference between node i and node j is calculated, that is, Δ ij =|L ij ∣, where L ij is the corresponding element between node i and node j in the Laplace matrix, indicating the degree of dependence on the power flow path; for node i, its total dependence is obtained by accumulating the difference in dependence with all other nodes, that is, , where D i represents the dependency of node i, reflecting the degree to which node i is affected by other nodes in the power grid; then, based on the total dependency relationship D i , analyze the importance of each node in the power grid, identify key nodes and power flow paths, and the nodes with higher node dependencies will bear more power transmission or scheduling tasks in the power grid; according to the dependency relationship between nodes Δ ij , fitting the main path of power flow, and analyzing the total dependence D i , assign appropriate scheduling coefficients to different nodes to ensure the rationality of load distribution; then, according to the fitted scheduling path and scheduling coefficient, an adaptive scheduling module is constructed, which can dynamically adjust the power flow between nodes in the power grid and optimize the power transmission path. The input of the adaptive scheduling module is the scheduling target. The module will calculate the total dependency of each node through the scheduling target, and output the optimized scheduling path and the adjusted node power distribution to ensure the efficient and stable operation of the power grid.

[0049] In summary, the embodiments of the present application have at least the following technical effects:

[0050] The embodiment of the present application obtains weather change data based on short-cycle time granularity, and builds a similar day database in combination with similar day features and historical data; through multi-level time granularity and rolling prediction technology, feature aggregation and data processing are performed to build a stacked LSTM model to predict power source and load; based on the prediction results, the source-load balance relationship is evaluated, and an adaptive scheduling module is built by establishing the target power grid topology and node dependencies; this module optimizes the power grid scheduling information through source-load balance compensation and scheduling path analysis to ensure the balance of power load and supply, and improve the stability and efficiency of power grid operation. These technical effects jointly solve the technical problem of inaccuracy in power source and load prediction caused by short-cycle drastic weather changes, and achieve the effect of improving the accuracy of short-cycle new energy prediction and short-term load prediction by combining similar day features and stacked LSTM models.

[0051] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. The processes depicted in the accompanying drawings do not necessarily require the specific order and continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0052] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0053] This specification and drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. A power grid adaptive dispatching method combined with short-cycle power source and load prediction, characterized in that: include: Acquire weather change information according to a short-period preset time granularity, perform feature extraction on the weather change information, and obtain weather change features, wherein the short-period preset time granularity includes multiple levels of time granularity; According to the short-period preset time granularity, weather change characteristics and current time positioning, similar day features are constructed, and historical data are extracted based on the similar day features to construct a similar day database; Preprocessing the similar day database, and performing feature aggregation on the similar day database to obtain similar day aggregate features, wherein the similar day aggregate features have a time granularity identifier; Inputting the weather change characteristics into a power grid prediction model to obtain a short-term power source and load prediction result, wherein the power grid prediction model is a stacked LSTM model, which is learned through a training data set constructed by similar daily aggregation features; Performing adaptive grid dispatching according to the balance relationship of the short-cycle power source and load prediction results; Preprocessing the similar day database and performing feature aggregation on the similar day database to obtain similar day aggregate features includes: Clean and filter the similar day database, construct a time window according to a preset time granularity, perform multi-level granularity data extraction on the similar day database, and construct a multi-level data architecture, where each level corresponds to a granular time window; Based on the multi-level data architecture, rolling prediction technology is used in the same level to identify feature relationships in granular time windows and perform horizontal feature rolling aggregation; Based on the multi-level data architecture, vertical feature aggregation is performed on the horizontal rolling aggregation features of the level granularity window according to the level neighborhood relationship to obtain the similar day aggregation features.

2. The grid adaptive dispatching method combined with short-cycle power source and load prediction according to claim 1, characterized in that: At the same level, rolling prediction technology is used to identify feature relationships in granular time windows and perform horizontal feature rolling aggregation, including: Based on the granular time window, target feature relationship recognition is performed to determine the relationship overlap rate; Identify the temporal relationship based on the granular time window and determine the time distance; Perform relationship fusion according to the relationship overlap rate and the time distance, and configure aggregation enhancement weights. The relationship overlap rate refers to the degree of feature similarity between two time windows. The more similar the features are, the higher the overlap rate is. The features in the granularity time window at the same level are weightedly aggregated according to the aggregation enhancement weight to obtain a horizontal scrolling aggregation feature.

3. The grid adaptive dispatching method combined with short-cycle power source and load prediction according to claim 1, characterized in that: Based on the multi-level data architecture, vertical feature aggregation is performed on the horizontal rolling aggregation features of the level granularity window according to the level neighborhood relationship, including: Identify the timing alignment relationship of hierarchical granularity windows; Based on the timing alignment relationship, hierarchical granularity window conversion mapping is performed to establish a timing alignment mapping relationship for each level of the data architecture; According to the time series alignment mapping relationship, the average values ​​of adjacent features are aggregated from small to large hierarchical granularity windows to obtain similar day aggregate features.

4. The grid adaptive dispatching method combined with short-cycle power source and load prediction according to claim 1, characterized in that: Inputting the weather change characteristics into the power grid prediction model also includes: Build a multi-layer stacked LSTM model architecture, including input layer, multiple LSTM layers, fully connected layer, and output layer; Based on the multi-level time granularity, construct a multi-level training data set and a verification data set according to the similar day aggregation features; The multi-level training data set and the verification data set are respectively used to perform multiple LSTM layer training and verify convergence to obtain the power grid prediction model.

5. The grid adaptive dispatching method combined with short-cycle power source and load prediction according to claim 4 is characterized in that: Based on the multi-level time granularity, a multi-level training data set and a verification data set are constructed according to the similar day aggregation features, including: Using the multi-level time granularity as an index, performing a time granularity identifier matching search in the similar day aggregation features to obtain similar day aggregation features of each time granularity; According to the prediction target, target identification is performed on the similar day aggregation features to establish a standardized training data set; Perform data formatting based on the standardized training data set and convert it into a format that conforms to the LSTM model input format, including the number of samples, time steps, and number of features; According to the time series of the similar day aggregate features, the formatted training data set is divided into a training data set and a verification data set, wherein the time series of the training data set is earlier than the time series of the verification data set.

6. The grid adaptive dispatching method combined with short-cycle power source and load prediction according to claim 1, characterized in that: According to the balance relationship of the short-cycle power source and load prediction results, the power grid is adaptively dispatched, including: According to the short-term power source and load forecast results, obtaining power load forecast results and energy output forecast results; Evaluate the source-load balance relationship according to the power load forecast result and the energy output forecast result to obtain the source-load balance relationship; Establish the target power grid topology, including power generation energy nodes and load center nodes; Identify and extract node dependencies according to the target power grid topology structure, and build an adaptive scheduling module; Performing balance compensation according to the source-load balance relationship to obtain a scheduling target; The scheduling target is used as input, and the scheduling path is parsed through the adaptive scheduling module to obtain adaptive scheduling information, where the adaptive scheduling information includes scheduling nodes and scheduling adjustment amounts thereof.

7. The grid adaptive dispatching method combined with short-cycle power source and load prediction according to claim 6, characterized in that: According to the target power grid topology structure, node dependency is identified and extracted, and an adaptive scheduling module is constructed, including: Based on the target power grid topology, the connection relationship between nodes is obtained to construct an adjacency matrix; Based on the target power grid topology, obtaining the number of adjacent nodes of the node and constructing a degree matrix; Performing difference calculation according to the adjacency matrix and the degree matrix to obtain dependency relationships of each node; According to the dependency relationship of each node, the balanced identification relationship of the scheduling path and the scheduling coefficient is fitted to construct the adaptive scheduling module.

Images