Multi-level power load supplementary prediction device based on multi-scale data features

Abstract

The invention provides a multi-level power load supplement prediction device based on multi-scale data features, and the device comprises a data integration module which is used for collecting and formatting multi-source data; the time multi-scale processing module is used for improving the frequency of low-time-frequency data in the formatted data so as to obtain data with the same frequency; the data collection module is used for combining the data with the same frequency into a format required by the multi-level scale deep learning model, performing hierarchical scale classification on each data feature by using a configuration file setting mode, and naming and distinguishing classification results through different data columns, and is used for constructing the multi-level scale deep learning model; finally, performing independent normalization processing on each column of data; and the multi-level scale deep learning model module is used for carrying out multi-level scale deep learning on the normalization processing to output a predicted value.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a multi-level electric power load supplement prediction device based on multi-scale data features. Background Art

[0002] Power load forecasting is an indispensable part of power market transactions. It is used to predict power demand in the next 24 hours, 48 ​​hours or even longer. In terms of forecast types, it can also be divided into regional power load forecasts based on administrative divisions such as provinces, regions, and cities, or power load forecasts based on electricity users such as buildings, factories, and parks.

[0003] For example, compared with the patent content with publication number CN202210238436.9, its method mainly collects data from power grid nodes, trains and predicts through CNN+LSTM model, and uses quadratic programming to optimize node data to meet the needs of power grid levels. From the content, this method does not specifically consider collecting data of different time frequencies, and this method uses CNN+LSTM model independently for each node, and does not aggregate independent node prediction data into higher-level prediction data.

[0004] Compared with the patent content with publication number CN117972410A, its method mainly utilizes an adaptive spatiotemporal synchronous graph convolutional neural network. On the basis of clustering the source data using the K-means method, the data is trained to make three data predictions: load prediction, wind power prediction, and photovoltaic prediction. Then, the three parts of the prediction results are optimized and corrected through the minimum trace level prediction method, thereby improving the prediction accuracy.

[0005] Compared with the patent content with publication number CN117977587B, it uses a multi-scale convolutional network to directly process the data in the feature engineering after collecting the data, and does not use other methods for data processing. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a multi-level power load supplement prediction device based on multi-scale data features in view of the above-mentioned defects in the prior art.

[0007] According to the present invention, a multi-level power load supplement prediction device based on multi-scale data features is provided, comprising: Data integration module, used for the collection and formatting of multi-source data; A time multi-scale processing module is used to increase the frequency of low time frequency data in the formatted data to obtain data with the same frequency; The data collection module is used to combine data with the same frequency into the format required by the multi-level scale deep learning model. Each data feature is classified into hierarchical scales using the configuration file settings. The classification results are distinguished by different data column names, which are used to build a multi-level scale deep learning model. Finally, each column of data is independently normalized. A multi-level scale deep learning model module is used to perform multi-level scale deep learning on the normalized processing to output the predicted value.

[0008] Preferably, the time multi-scale processing module performs data analysis on the data with the highest time frequency among different types, calculates data features of different low time frequencies from all historical data of the type to which the highest time frequency belongs, and then interpolates and reconstructs low time frequency data of the same type but obtained from other sources according to the calculated data features to form new high-frequency data.

[0009] Preferably, the structure of the multi-level scale deep learning model is designed as an inverted tree structure, where each branch of the tree corresponds to a sub-prediction module in the actual scenario, and the combination is screened by the hierarchical labels in the data collection module, that is, each branch contains one or a group of data feature data obtained in the previous step.

[0010] Preferably, each branch contains an independent prediction model, and the data of the same level is aggregated through a weighted gate module; depending on whether the prediction target is a single data feature or a multivariate data feature, the input of each weighted gate is a number of one-dimensional weight vectors, the number of which is consistent with the number of branches under it, and the size of a single vector is the number of features to be predicted by its corresponding branch.

[0011] Preferably, the prediction model is a CNN+GRU+ATTENTION or WNN+CNN+ATTENTION combination model.

[0012] Preferably, the multi-level power load supplementary prediction device based on multi-scale data features also includes a supplementary prediction module, which is used to obtain prediction data by combining similar data with calculated data features according to the required prediction time frequency.

[0013] Preferably, the multi-level power load supplementary prediction device based on multi-scale data features also includes a data output module, which is used to combine the prediction results in the multi-level scale deep learning model module and the supplementary prediction data generated by the supplementary prediction module into new data, and input them into the deep learning network in the multi-level scale deep learning model module for training and prediction, so as to finally obtain updated prediction data.

[0014] Preferably, the final data is output after being denormalized by RevIN to obtain the corresponding load prediction value.

[0015] Therefore, the present invention proposes a multi-level power load supplementary forecasting method based on multi-scale data features. This method not only further considers the characteristics of different scales of various types of data involved in power load forecasting, but also proposes a multi-level model architecture scheme based on its common application scenarios, which can effectively improve the scenario adaptability and forecasting performance. Experiments have found that this method can effectively utilize the meteorological data information of prefecture-level cities under the jurisdiction of the target test province, as well as meteorological data information and load information of different time frequencies; and the prediction results are consistent with the actual results at a high level. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] A more complete understanding of the present invention and its attendant advantages and features will be more readily appreciated by reference to the following detailed description taken in conjunction with the accompanying drawings, in which: Figure 1 The functional block diagram of a multi-level power load replenishment prediction device based on multi-scale data features according to a preferred embodiment of the present invention is schematically shown.

[0017] Figure 2 The structure of the multi-level scale deep learning model module of the multi-level power load replenishment prediction device based on multi-scale data features according to the preferred embodiment of the present invention is schematically shown.

[0018] It should be noted that the drawings are used to illustrate the present invention, rather than to limit the present invention. Note that the drawings showing the structures may not be drawn to scale. In addition, in the drawings, the same or similar elements are marked with the same or similar reference numerals. DETAILED DESCRIPTION

[0019] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0021] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0022] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] On the basis of the prior art, the present invention proposes the following technical solution.

[0024] (1) The present invention optimizes the content of the data processing / data engineering module of the prior art, and allows two types of multi-scale features to exist in the data, namely, time multi-scale and data level scale, depending on the usage scenario. Time multi-scale means that data with different time frequencies are allowed to exist in the data, such as daily frequency, hour frequency, minute frequency, etc.; in addition, data with different levels of granularity are allowed to exist in the data. For example, in the scenario of power load forecasting for a certain province, data related to prefecture-level cities or finer-grained county-level cities are allowed to exist in the data; in the scenario of power load forecasting for a park, data related to each building in the park is allowed to exist in the data, etc. This increases the flexibility of model migration in multiple scenarios.

[0025] (2) The present invention optimizes the multi-scale data alignment method / module of the prior art. For data of different time scales, the high time frequency data is first analyzed to obtain basic information such as data distribution, maximum value, mean, variance, etc.; then the same type of low time frequency data is converted from low frequency to high frequency according to the time frequency to obtain new data. The new data is then processed by the data smoothing module to eliminate data mutations and abnormal differences in data at adjacent time points. This solution can manually retain the high-frequency useful information features of historical data, thereby improving the prediction accuracy of the model.

[0026] (3) The present invention optimizes the deep neural network structure of the prior art and proposes a hierarchical construction optimization scheme based on practical scenarios. It mainly performs learnable weighted aggregation on low-level data and uses the aggregation results for training and prediction of higher-level models. This scheme is more suitable for the simulation of the targeted scenarios and improves scenario adaptability and prediction accuracy.

[0027] (4) The present invention adaptively optimizes the model training method of the prior art, combines the prediction method of the residual network, and designs a residual supplementary prediction scheme based on the fusion of multivariate data such as similar data, third-party data, and prediction data. Such a design can better utilize the data features of different source data, thereby increasing the possibility of retaining local data features in the original data in the predicted data.

[0028] Specifically, Figure 1 The functional block diagram of the multi-level power load supplement prediction device based on multi-scale data features according to a preferred embodiment of the present invention is schematically shown. Figure 1 As shown, the multi-level power load supplement prediction device based on multi-scale data features according to the preferred embodiment of the present invention includes: The data integration module 10 is used for collecting and formatting multi-source data; for example, the collected data includes power load data, meteorological data, calendar data, electricity price policy data, and power coal price index data. The collected data is formatted to form formatted data entries from earliest to latest.

[0029] The time multi-scale processing module 20 is used to increase the frequency of low time frequency data in the formatted data to obtain data with the same frequency (ie, data with all time scales aligned).

[0030] There are many ways to increase the frequency. In a preferred example, for example, first, the time multi-scale processing module performs data analysis on the data with the highest time frequency in different types, and calculates different data features of low time frequencies from all historical data of the type to which the highest time frequency belongs. For example, if the highest frequency data is minute frequency data, and the lower frequency data is hour frequency, then the module will calculate the mean, variance, maximum, percentile and other data features from all minute frequency data corresponding to each hour frequency. For example, if the minute data is data in the interval from 1:00 to 2:00 and the interval is 10 minutes, then it includes data at six time points of 1:10, 1:20, 1:30, 1:40, 1:50, and 2:00, then the mean, variance, maximum, minimum, percentile and other data can be calculated for the data at these six points as data features.

[0031] After completing the calculation, the module will interpolate and reconstruct low-frequency data of the same type but obtained from other sources (for example, meteorological data obtained from different data sources, some of which are hourly, some are three-hourly or daily) according to the calculated data features to form new high-frequency data.

[0032] The data collection module 30 is used to combine data with the same frequency into the format required by the multi-level scale deep learning model, classify each data feature into levels using the configuration file settings, and distinguish the classification results by naming different data columns, which are used to build a multi-level scale deep learning model, and finally perform independent normalization on each column of data.

[0033] Among them, the hierarchical scale classification will have different classification results according to different usage scenarios. For example, hierarchical scale classification for a certain park can be carried out to classify different types of power loads in the park, such as office buildings, production workshops, etc., and further, different types of load equipment in office buildings can be classified, such as air conditioning equipment, computers, etc. Finally, the smallest classification includes all data involved in the classification. Each sub-classification will generate a corresponding hierarchical label.

[0034] The multi-level scale deep learning model module 40 is used to perform multi-level scale deep learning on the normalization process to output a prediction value.

[0035] Preferably, if Figure 2 As shown in the figure, the structural design of the multi-level scale deep learning model is mainly an inverted tree structure. Each branch of the tree corresponds to a sub-prediction module in the actual scenario, and the combination is screened by the hierarchical labels in the data collection module, that is, each branch contains one or a group of data feature data obtained in the previous step; each branch contains an independent prediction model, and the data at the same level is aggregated through a weighted gate module (used to perform weighted operations on the results generated by the prediction models of the branches at the same layer); depending on whether the prediction target is a single data feature or a multivariate data feature, the input of each weighted gate is a one-dimensional weight vector, and the number of one-dimensional weight vectors is consistent with the number of branches under it, and the size of a single one-dimensional weight vector is the number of features to be predicted by its corresponding branch.

[0036] The prediction model can adopt various machine learning and deep learning prediction models used in the industry. In the practice of this method, the CNN+GRU+ATTENTION or WNN+CNN+ATTENTION combination model is preferred.

[0037] For example, the present invention modifies the calculation method of weighting the weighted vector, uses a fully connected network to align, transpose and multiply the weight vector and the output result of the CNN+GRU+ATTENTION or WNN+CNN+ATTENTION model, obtains the weighted result, and then reduces the dimension of the output through the fully connected layer. The root of this hierarchical tree structure outputs the predicted value.

[0038] The supplementary prediction module 50 is used to obtain prediction data by combining similar data by calculating data features according to the time frequency required for prediction.

[0039] Preferably, the supplementary prediction module uses cosine similarity or Euclidean distance to calculate data feature similarity data. If third-party prediction data (target prediction data obtained from other data sources) can be obtained in the data integration module 10, it is used as a type of prediction data. In addition, the predicted output value of the multi-level scale deep learning model module 40 is considered. The difference between the historical prediction value and these three types of data (the predicted output value of the multi-level scale deep learning model module 40, the predicted data obtained by calculating the data feature similarity, and the target prediction data obtained from other data sources) can be calculated respectively. By reusing the weighted gate module and the prediction model described in the multi-level scale deep learning model module 40, the three types of differences can be predicted and aggregated to obtain supplementary prediction data.

[0040] The data output module 60 is used to form new data from the prediction results in the multi-level scale deep learning model module 40 and the supplementary prediction data generated by the supplementary prediction module 50, and input them into the deep learning network in the multi-level scale deep learning model module 40 for training and prediction, and finally obtain updated prediction data. Preferably, the final data is output after RevIN denormalization processing to obtain the corresponding load prediction value.

[0041] In addition, it should be noted that, unless otherwise specified, the terms "first", "second", "third", etc. in the specification are only used to distinguish the various components, elements, steps, etc. in the specification, and are not used to indicate the logical relationship or sequential relationship between the various components, elements, steps, etc.

[0042] Therefore, the present invention proposes a multi-level power load supplement forecasting method based on multi-scale data features. This method not only further considers the characteristics of different scales of various types of data involved in power load forecasting, but also proposes a multi-level model architecture scheme according to its common application scenarios, which can effectively improve the scenario adaptability and prediction performance.

[0043] It is to be understood that, although the present invention has been disclosed as a preferred embodiment, the above embodiment is not intended to limit the present invention. For any person skilled in the art, without departing from the scope of the technical solution of the present invention, the technical content disclosed above can be used to make many possible changes and modifications to the technical solution of the present invention, or modified into equivalent embodiments of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A multi-level power load supplement forecasting device based on multi-scale data features, characterized in that include: Data integration module, used for the collection and formatting of multi-source data; A time multi-scale processing module is used to increase the frequency of low time frequency data in the formatted data to obtain data with the same frequency; The data collection module is used to combine data with the same frequency into the format required by the multi-level scale deep learning model. Each data feature is classified into hierarchical scales using the configuration file settings. The classification results are distinguished by different data column names, which are used to build a multi-level scale deep learning model. Finally, each column of data is independently normalized. A multi-level scale deep learning model module, used to perform multi-level scale deep learning on the normalization process to output a predicted value; Among them, the structure of the multi-level scale deep learning model is designed as an inverted tree structure. Each branch of the tree corresponds to a sub-prediction module in the actual scenario, and the combination is screened by the hierarchical labels in the data collection module, that is, each branch contains one or a group of data feature data obtained in the previous step; each branch contains an independent prediction model, and the data at the same level are aggregated through a weighted gate module; the input of each weighted gate is a one-dimensional weight vector, and the number of one-dimensional weight vectors is consistent with the number of branches under it, and the size of a single one-dimensional weight vector is the number of features to be predicted by its corresponding branch.

2. The multi-level power load supplement prediction device based on multi-scale data features according to claim 1 is characterized in that: The time multi-scale processing module performs data analysis on the data with the highest time frequency in different types, and calculates different data features of low time frequencies from all historical data of the type with the highest time frequency. Subsequently, the low time frequency data of the same type but obtained from other sources are interpolated and reconstructed according to the calculated data features to form new high-frequency data.

3. The multi-level power load supplement prediction device based on multi-scale data features according to claim 1 or 2, characterized in that: The prediction model is a CNN+GRU+ATTENTION or WNN+CNN+ATTENTION combination model.

4. The multi-level power load supplement prediction device based on multi-scale data features according to claim 1 or 2 is characterized in that It also includes a supplementary prediction module, which is used to obtain prediction data by combining similar data by calculating data features according to the time frequency required for prediction.

5. The multi-level power load supplement prediction device based on multi-scale data features according to claim 1 or 2 is characterized in that It also includes a data output module, which is used to combine the prediction results in the multi-level scale deep learning model module and the supplementary prediction data generated by the supplementary prediction module into new data, which is input into the deep learning network in the multi-level scale deep learning model module for training and prediction, and finally obtains updated prediction data.

6. The multi-level power load supplement prediction device based on multi-scale data features according to claim 1 or 2, characterized in that: The final data is output after being denormalized by RevIN to obtain the corresponding load forecast value.

Images