Ultra-short-term load forecasting method and system based on two-stage intelligent feature engineering

Abstract

The present disclosure belongs to the technical field of power systems, and specifically relates to a method and system for ultra-short-term load forecasting based on two-stage intelligent feature engineering, comprising: acquiring historical data of power load; performing cluster analysis on the acquired historical data of power load to obtain an optimal historical daily load curve; and completing ultra-short-term forecasting of power load based on the obtained optimal historical daily load curve and a preset forecasting model; wherein the preset forecasting model adopts a combined forecasting model, constructs an ultra-short-term forecasting model through an extreme learning machine, and combines the Bayesian optimization extreme learning machine to achieve ultra-short-term forecasting of power load.

Description

Technical Field

[0001] The present disclosure belongs to the technical field of power systems, and in particular relates to a method and system for ultra-short-term load forecasting based on two-stage intelligent feature engineering. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] As one of the most important energy sources, electricity plays a key role in global economic development and human production and life. Demand-side response is a crucial component of smart grid development and building energy conservation and emission reduction. It is a short-term approach aimed at guiding users to use electricity rationally, reducing peak and valley load variations, and conserving energy. It relies on accurate power load forecasting; high-precision power load forecasting is crucial for developing grid dispatch plans and helping users increase efficiency and reduce costs.

[0004] According to the inventors' understanding, there are three main types of power load prediction models:

[0005] The first type is a physics-based model that uses physical information such as atmospheric pressure, temperature, and material, as well as a large number of physical mechanisms, to derive a set of mathematical equations for prediction. However, physics-based models struggle to find appropriate equations to describe the complex nonlinear relationships between loads.

[0006] The second category is models based on statistics. Currently, the most widely used models include regression models, moving average models, autoregressive moving average models, autoregressive integrated moving average models, and seasonal autoregressive integrated moving average models. However, these models have poor robustness and poor prediction performance for complex nonlinear and non-stationary series.

[0007] The third category is machine learning models, such as Extreme Learning Machine (ELM), Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Long-Short Term Memory (LSTM). Machine learning models have good performance in processing nonlinear time series problems.

[0008] Existing technologies use feature engineering methods to process historical load data and environmental characteristics, improving the predictive capabilities of machine learning models. However, these methods only analyze the impact of individual weather information such as temperature and humidity on load forecasts, or simply use numerous external factors as model features. They lack feature engineering that rationally and efficiently combines multi-dimensional influencing factors, making it impossible to balance the stability of load trend forecasts and the accuracy of fluctuation detail forecasts. Summary of the Invention

[0009] In order to solve the above problems, the present disclosure proposes a load ultra-short-term forecasting method and system based on two-stage intelligent feature engineering, which efficiently utilizes multi-dimensional influencing factors and can balance the stability of load trend forecasting and the accuracy of fluctuation detail prediction.

[0010] According to some embodiments, a first solution of the present disclosure provides an ultra-short-term load forecasting method based on two-stage intelligent feature engineering, which adopts the following technical solutions:

[0011] A method for ultra-short-term load forecasting based on two-stage intelligent feature engineering, including:

[0012] Obtain historical data on power load;

[0013] Perform cluster analysis on the acquired historical power load data to obtain the optimal historical daily load curve;

[0014] Based on the obtained optimal historical daily load curve and the preset forecasting model, the ultra-short-term forecast of power load is completed;

[0015] Among them, the preset prediction model adopts a combined prediction model, constructs an ultra-short-term prediction model through the extreme learning machine, and combines the Bayesian optimization extreme learning machine to realize the ultra-short-term prediction of power load.

[0016] As a further technical limitation, the multi-dimensional influencing factors of the acquired historical data on power load include weather effects, calendar effects and historical correlation of users' electricity consumption behaviors.

[0017] Furthermore, before obtaining the optimal historical daily load curve, similar historical days are selected by considering the impact of calendar effect and weather effect on future load trend changes. K-Means++ is used to perform cluster analysis on the historical power load, and the Elbow method is used to determine the k value to complete the cluster analysis of the acquired historical power load data.

[0018] As a further technical limitation, the power load sequence in the optimal historical daily load curve is decomposed into power load cycle component, power load trend component and power load residual component through time series decomposition method to obtain fine-grained data; the convolutional neural network and long short-term memory neural network are used to learn the spatiotemporal characteristics of the power load trend component to obtain the load trend forecast value for the next day.

[0019] Furthermore, in the process of obtaining the load trend forecast value for the next day, the power load sequence in the optimal historical daily load curve is trained through a hybrid deep learning model, and the output value of its output layer is the load trend forecast value for the next day; the hybrid deep learning model includes a time series decomposition module, a convolutional neural network spatial feature extraction module, a long short-term memory neural network time feature extraction module and an output layer.

[0020] As a further technical limitation, in the process of constructing an ultra-short-term prediction model through an extreme learning machine, a power consumption behavior correlation matrix is ​​constructed based on the acquired power load historical data, the user power consumption behavior correlation is calculated, and the correlation of the user power consumption behavior within a specified time period is obtained; by judging the size between the obtained correlation and the correlation threshold, the training set is updated to complete the construction of the ultra-short-term prediction model.

[0021] As a further technical limitation, the Bayesian optimization extreme learning machine includes selection of parameters to be optimized, parameter initialization, parameter optimization and optimal parameter feedback.

[0022] According to some embodiments, a second solution of the present disclosure provides an ultra-short-term load forecasting system based on two-stage intelligent feature engineering, which adopts the following technical solutions:

[0023] A load ultra-short-term forecasting system based on two-stage intelligent feature engineering, including:

[0024] an acquisition module configured to acquire historical data of power load;

[0025] an analysis module configured to perform cluster analysis on the acquired historical power load data to obtain an optimal historical daily load curve;

[0026] A prediction module is configured to perform ultra-short-term prediction of power load based on the obtained optimal historical daily load curve and a preset prediction model;

[0027] Among them, the preset prediction model adopts a combined prediction model, constructs an ultra-short-term prediction model through the extreme learning machine, and combines the Bayesian optimization extreme learning machine to realize the ultra-short-term prediction of power load.

[0028] According to some embodiments, a third solution of the present disclosure provides a computer-readable storage medium, which adopts the following technical solution:

[0029] A computer-readable storage medium stores a program thereon, which, when executed by a processor, implements the steps in the ultra-short-term load forecasting method based on two-stage intelligent feature engineering as described in the first aspect of the present disclosure.

[0030] According to some embodiments, a fourth solution of the present disclosure provides an electronic device, which adopts the following technical solution:

[0031] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the ultra-short-term load forecasting method based on two-stage intelligent feature engineering as described in the first aspect of the present disclosure are implemented.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The present disclosure provides an ultra-short-term prediction method for power load based on two-stage intelligent feature engineering. Through the two-stage intelligent feature engineering, the multi-dimensional influencing factors in the complex power environment (including weather effects, calendar effects and historical correlation of users' electricity consumption behaviors) are reasonably and fully considered to eliminate redundant features in historical data, thereby improving the learning efficiency and prediction accuracy of the model. The combined model structure is used to simultaneously consider trend stability and fluctuation detail prediction accuracy, thereby greatly improving the prediction accuracy and adaptability of the prediction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings, which constitute a part of the present disclosure, are used to provide a further understanding of the present disclosure. The exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation to the present disclosure.

[0035] Figure 1 This is a flow chart of the ultra-short-term load forecasting method based on two-stage intelligent feature engineering in the first embodiment of the present disclosure;

[0036] Figure 2 This is a structural diagram of the ultra-short-term load forecasting method based on two-stage intelligent feature engineering in the first embodiment of the present disclosure;

[0037] Figure 3 Schematic diagram of the STL-CLSTM model structure in the first embodiment of the present disclosure;

[0038] Figure 4 This is a flowchart of the ELM ultra-short-term load forecasting considering the historical correlation of user electricity consumption behavior in the first embodiment of the present disclosure;

[0039] Figure 5 This is a structural block diagram of the ultra-short-term load forecasting system based on two-stage intelligent feature engineering in the second embodiment of the present disclosure. DETAILED DESCRIPTION

[0040] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0041] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present disclosure belongs.

[0042] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0043] In the present disclosure, terms such as "upper", "lower", "left", "right", "front", "back", "vertical", "horizontal", "side", "bottom", etc. indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. They are merely relational words determined for the convenience of describing the structural relationships of the various parts or elements of the present disclosure, and do not specifically refer to any part or element in the present disclosure, and should not be understood as limitations on the present disclosure.

[0044] In this disclosure, terms such as "fixed connection," "connected," and "connection" should be interpreted broadly to mean a fixed connection, an integral connection, or a detachable connection; a direct connection or an indirect connection through an intermediary. Relevant researchers or technicians in this field may determine the specific meanings of these terms in this disclosure based on specific circumstances, and they should not be construed as limitations on this disclosure.

[0045] In the absence of conflict, the embodiments of the present disclosure and the features thereof may be combined with each other.

[0046] Example 1

[0047] The first embodiment of the present disclosure introduces a method for ultra-short-term load forecasting based on two-stage intelligent feature engineering.

[0048] like Figure 1 The ultra-short-term load forecasting method based on two-stage intelligent feature engineering is shown, including:

[0049] Obtain historical data on power load;

[0050] Perform cluster analysis on the acquired historical power load data to obtain the optimal historical daily load curve;

[0051] Based on the obtained optimal historical daily load curve and the preset forecasting model, the ultra-short-term forecast of power load is completed;

[0052] Among them, the preset prediction model adopts a combined prediction model, constructs an ultra-short-term prediction model through the extreme learning machine, and combines the Bayesian optimization extreme learning machine to realize the ultra-short-term prediction of power load.

[0053] This embodiment proposes an ultra-short-term power load forecasting method based on two-stage intelligent feature engineering. This method fully considers various external factors, such as temperature, humidity, and calendar, as well as the random user behavior of electricity consumption over short periods of time. This removes redundant features from the load sequence, helping to improve the model's learning efficiency and forecasting accuracy. Furthermore, the present invention establishes a combined model structure that not only fully integrates the advantages of each algorithm, but also balances trend stability and the accuracy of forecasting fluctuation details.

[0054] like Figure 2 As shown in the figure, this embodiment is based on the ultra-short-term load forecasting model of STL, CNN, LSTM and ELM, and introduces the hyperparameters of the BOA optimization model; the first phase of intelligent feature engineering is carried out to obtain a stable load trend for the next day: the Spearman's rank correlation coefficient (Spearman's rank correlation coefficient) is used to calculate the load trend of the next day. The comprehensive correlation between multi-dimensional influencing factors and power load is calculated using the correlation coefficient. Based on this, the best similar daily load curve is searched for in combination with numerical weather forecasts and calendar information. The best similar daily load curve is decomposed using the advantage of STL in extracting dynamic features of time series to obtain trend components. The trend components are processed using the powerful spatiotemporal feature extraction capabilities of CNN and LSTM to obtain trend forecasts for the next day. Based on these trend forecast results, the second phase of intelligent feature engineering is carried out to process the details of user electricity usage behavior and remove redundant features in the power load series, which helps improve model learning efficiency. The fast-learning ELM is used to achieve ultra-short-term forecasts, and the BOA is introduced to optimize model hyperparameters. Specifically, guided by the results of the first phase of intelligent feature engineering (day-ahead load trends), time windows are set, and the correlation of historical loads in each time window is dynamically calculated. Data with low correlation is eliminated (referred to as redundant feature removal) to obtain a dynamic training set. The ELM is dynamically trained based on this dynamic training set, and the load fluctuation details are predicted. To accelerate prediction, the ELM is selected and combined with the BOA to optimize the ELM hyperparameters.

[0055] As one or more implementations, input variables for the prediction model are determined. Specifically, 31 consecutive days of historical user power load data are collected with a data interval of 15 minutes and divided into a training set (25 days) and a test set (6 days). The training set is used to obtain the hybrid prediction model, and the test set is used to verify the prediction accuracy of the method.

[0056] In the first stage of intelligent feature engineering in this embodiment, the optimal similar daily load curve is searched, and then the load trend forecast value for the next day is obtained based on the obtained optimal similar daily load curve.

[0057] When selecting historically similar days in the first phase of intelligent feature engineering, the impact of calendar and weather effects on future load trends was taken into account. Calendar effects can not only reflect the basic trends of power load changes, but also, to a certain extent, reflect the basic laws of user electricity consumption behavior.

[0058] This example uses K-Means++ to cluster historical power loads and uses the Elbow method to determine the k value. Power loads are classified according to calendar type into weekday loads, weekend loads, and other holiday loads. Weather is a key factor influencing user electricity usage. For example, outdoor temperature affects the operating status of a building's air conditioning system, and solar irradiance affects the operating status of a building's lighting system.

[0059] Therefore, in addition to the calendar effect, it is also necessary to consider the impact of weather effects on power load, which includes the following three steps:

[0060] ①Use the normalization function shown in formula (1) to normalize all numerical weather forecast (including temperature, humidity, solar irradiance, wind speed) and power load data;

[0061]

[0062] Among them, Y represents the original signal, Y max and Y min Represent the maximum and minimum values ​​of the original signal, Y norm is the signal after normalization.

[0063] ② The Spearman rank correlation coefficient between power load and numerical weather forecast (including temperature, humidity, solar irradiance, wind speed) is calculated by formula (2), which is denoted as t w ,h w ,s w ,w w ; t w ,h w ,s w ,w wSubstituting into formula (3) we can get the impact of numerical weather forecast on future load changes, which is denoted as T w ,H w ,S w ,W w . Where d i is the different level between two data variables, and n is the number of data.

[0064]

[0065]

[0066] ③ Use formula (2) to calculate the correlation between the numerical weather forecast of each day in history and the numerical weather forecast of the forecast day, which are recorded as Ti, Hi, Si, and Wi respectively; calculate the comprehensive similarity coefficient SL i ., the calculation process is shown in formula (4): SL i =T w ·T i +H w ·H i +S w ·S i +W w W i (4)

[0067] The daily load curve on the date with the maximum comprehensive similarity coefficient is regarded as the optimal daily load curve.

[0068] In the process of obtaining the load trend forecast value for the next day, STL is used to decompose the power load series into periodic component, trend component and residual component, so as to obtain finer-grained data and reduce the impact of irregular fluctuations on prediction accuracy; among them, the trend component is the result of smoothing the original signal, and its volatility is greatly reduced, so the prediction difficulty is reduced accordingly.

[0069] By combining CNN and LSTM to construct an STL-CLSTM hybrid deep learning model, the STL-CLSTM hybrid deep learning model is used to learn the spatiotemporal characteristics of the power load trend component and obtain the load trend forecast value for the next day.

[0070] In this embodiment, the STL-CLSTM hybrid deep learning model includes four parts: time series decomposition module, CNN spatial feature extraction module, LSTM temporal feature extraction module and output layer. Figure 3 As shown in the figure, the CNN spatial feature extraction module consists of Conv layer, Flatten and Dropout. The calculation formula of one-dimensional convolution is shown in formula (5), which is used to extract data spatial features; that is,

[0071]

[0072] Among them, a j (l+1) (τ) represents the feature map j of layer l+1, σ represents the nonlinear activation function, F l represents the number of feature maps in layer l, The representative kernel performs convolution processing on the feature map f of the lth layer to form the feature map j of the l+1th layer, represents a deviation vector, p l Represents the kernel length in layer l.

[0073] The Flatten layer controls data dimensionality; the Dropout layer is added to prevent overfitting during training. The LSTM temporal feature extraction module consists of two LSTM layers and one Dense layer. As a special type of RNN, the LSTM is composed of units called memory blocks. Each memory block effectively addresses the exploding and vanishing gradient problems of RNNs by adding three gates: an input gate, a forget gate, and an output gate. Finally, the output information is converted into power load forecasts through the fully connected layer of the output layer.

[0074] During the second phase of intelligent feature engineering in this embodiment, load trends reflect the characteristics of user electricity usage behavior. Historically, the similarity of load curves within the same time period but on different days reflects the historical correlation between users' electricity usage behaviors during that time period. Stronger correlations indicate more regular user electricity usage during that time period, while conversely, more uncertain user electricity usage behaviors during that time period.

[0075] Therefore, this embodiment quantifies the historical correlation of user electricity consumption behavior and incorporates it into the load forecasting framework, which can eliminate redundant data and improve the prediction performance of the model during high load fluctuation periods when user electricity consumption behavior is large. The specific process is as follows: Figure 4 As shown in the figure, by establishing a sliding window for electricity consumption behavior correlation analysis, establishing an electricity consumption behavior correlation matrix, and updating the training set, the ultra-short-term prediction of power load is achieved.

[0076] In this embodiment, the power load is collected at a frequency of 15 minutes per point, and a sliding window with a size of n and a step length of 15 minutes is used to calculate the correlation between user electricity consumption behaviors. The size of the sliding window for analyzing the correlation between electricity consumption behaviors has a great influence on the prediction accuracy: if the sliding window size is too large, the details of energy consumption behaviors will be ignored and more highly correlated behavioral features will be lost; conversely, if the sliding window size is too large, the prediction accuracy will be reduced due to excessive sensitivity to noise data. Therefore, the size should be selected based on the trial and error method.

[0077] The training set contains historical load data for D days, and the scale of the observation window is set to n and the step length is 15 minutes. The load at the kth point in the mth segment of the dth day is recorded as l k d,m , then the average load level on the dth day and the mth segment is L d,m , as shown in formula (6):

[0078]

[0079] Among them, d k is the difference between two data variables.

[0080] The Spearman rank correlation coefficient, Euclidean distance, behavior correlation and behavior correlation matrix of the load on day i and the load on day j in segment m are respectively denoted as ρ m i,j ,dist m i,j ,w m i,j and W m The correlation degree of the user's electricity consumption behavior in the mth segment is recorded as ω m , the calculation formulas are shown in formula (7), formula (8), formula (9), formula (10) and formula (11) respectively:

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] The behavior correlation degree ω in this embodiment m It can reflect whether the user's historical electricity consumption behavior has a strong correlation in the m period. If ω m If it is less than the threshold τ (determined by trial and error), it indicates that the historical correlation of the user's electricity usage behavior in this time period is poor and random. It is difficult to obtain high-precision prediction results by directly training the model with this part of data.

[0088] This embodiment uses the load trend in the next 24 hours as a guide for future load trend changes, and performs "purification" processing on the original data set to reduce the interference of power consumption behaviors with weak historical correlation on ultra-short-term prediction. According to formula (8), the behavioral correlation ρ between the i-th day (i belongs to [1, D]) in the original data set and the trend prediction result in the m-th sliding window is calculated. m i If ρ m i If the value is less than the threshold τ, it means that in the m period, the correlation between the electricity consumption behavior of the day and the electricity consumption behavior of the forecast day is very different. Using this part of the data to train the model will have a negative impact on the prediction accuracy of the model to a certain extent, and the degree of impact is related to the correlation degree ω of the m period. m Therefore, when predicting data for period m, the data for day i should be removed from the original data set. On the contrary, if ρ m i If ω is greater than b, the data of day i should be retained. m If it is greater than the threshold a, it means that the historical correlation of the user's energy consumption behavior in this period is strong and there is no need to process the training set.

[0089] This embodiment uses two-stage intelligent feature engineering to reasonably and fully consider the multi-dimensional influencing factors in the complex power environment (including weather effects, calendar effects and historical correlation of users' electricity consumption behaviors) and eliminate redundant features in historical data, thereby improving the learning efficiency and prediction accuracy of the model. The combined model structure is used to simultaneously consider trend stability and fluctuation detail prediction accuracy, which can greatly improve the prediction accuracy and adaptability of the prediction method.

[0090] Example 2

[0091] The second embodiment of the present disclosure introduces a load ultra-short-term forecasting system based on two-stage intelligent feature engineering.

[0092] like Figure 5 The ultra-short-term load forecasting system based on two-stage intelligent feature engineering is shown, including:

[0093] an acquisition module configured to acquire historical data of power load;

[0094] an analysis module configured to perform cluster analysis on the acquired historical power load data to obtain an optimal historical daily load curve;

[0095] A prediction module is configured to perform ultra-short-term prediction of power load based on the obtained optimal historical daily load curve and a preset prediction model;

[0096] Among them, the preset prediction model adopts a combined prediction model, constructs an ultra-short-term prediction model through the extreme learning machine, and combines the Bayesian optimization extreme learning machine to realize the ultra-short-term prediction of power load.

[0097] The detailed steps are the same as those of the ultra-short-term load forecasting method based on two-stage intelligent feature engineering provided in Example 1, and will not be repeated here.

[0098] Example 3

[0099] A third embodiment of the present disclosure provides a computer-readable storage medium.

[0100] A computer-readable storage medium stores a program, which, when executed by a processor, implements the steps of the ultra-short-term load forecasting method based on two-stage intelligent feature engineering as described in the first embodiment of the present disclosure.

[0101] The detailed steps are the same as those of the ultra-short-term load forecasting method based on two-stage intelligent feature engineering provided in Example 1, and will not be repeated here.

[0102] Example 4

[0103] A fourth embodiment of the present disclosure provides an electronic device.

[0104] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the ultra-short-term load forecasting method based on two-stage intelligent feature engineering as described in the first embodiment of the present disclosure are implemented.

[0105] The detailed steps are the same as those of the ultra-short-term load forecasting method based on two-stage intelligent feature engineering provided in Example 1, and will not be repeated here.

[0106] The foregoing description is merely a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present disclosure shall be included within the scope of protection of the present disclosure.

[0107] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Those skilled in the art should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. A method for ultra-short-term load forecasting based on two-stage intelligent feature engineering, characterized in that: include: Obtain historical data on power load; Perform cluster analysis on the acquired historical power load data to obtain the optimal historical daily load curve; Based on the obtained optimal historical daily load curve and the preset forecasting model, the ultra-short-term forecast of power load is completed; Among them, the preset prediction model adopts a combined prediction model, constructs an ultra-short-term prediction model through the extreme learning machine, and combines the Bayesian optimization extreme learning machine to achieve ultra-short-term prediction of power load; The power load sequence in the optimal historical daily load curve is decomposed into power load cycle components, power load trend components, and power load residual components through the time series decomposition method to obtain fine-grained data. The spatiotemporal characteristics of the power load trend components are learned using convolutional neural networks and long short-term memory neural networks to obtain the load trend forecast value for the next day. In the process of obtaining the load trend forecast value for the next day, the power load sequence in the optimal historical daily load curve is trained through a hybrid deep learning model, and the output value of its output layer is the load trend forecast value for the next day; the hybrid deep learning model includes a time series decomposition module, a convolutional neural network spatial feature extraction module, a long short-term memory neural network time feature extraction module and an output layer.

2. The method for ultra-short-term load forecasting based on two-stage intelligent feature engineering as described in claim 1, characterized in that: The multi-dimensional influencing factors of the acquired historical power load data include weather effects, calendar effects and historical correlation of users' electricity consumption behaviors.

3. The method for ultra-short-term load forecasting based on two-stage intelligent feature engineering as described in claim 2, characterized in that: Before obtaining the optimal historical daily load curve, similar historical days are selected by considering the impact of calendar effect and weather effect on future load trend changes. K-Means++ is used to perform cluster analysis on the historical power load, and the Elbow method is used to determine the k value to complete the cluster analysis of the acquired historical power load data.

4. The method for ultra-short-term load forecasting based on two-stage intelligent feature engineering as described in claim 1, characterized in that: In the process of building an ultra-short-term prediction model through extreme learning machines, a power consumption behavior correlation matrix is ​​constructed based on the acquired power load historical data, and the correlation of user power consumption behavior is calculated to obtain the correlation of user power consumption behavior within a specified time period; By judging the size between the obtained correlation and the correlation threshold, the training set is updated to complete the construction of the ultra-short-term prediction model.

5. The method for ultra-short-term load forecasting based on two-stage intelligent feature engineering as described in claim 1, characterized in that: The Bayesian optimization extreme learning machine includes the selection of parameters to be optimized, parameter initialization, parameter optimization and optimal parameter feedback.

6. A load ultra-short-term forecasting system based on two-stage intelligent feature engineering, characterized by: include: an acquisition module configured to acquire historical data of power load; an analysis module configured to perform cluster analysis on the acquired historical power load data to obtain an optimal historical daily load curve; A prediction module is configured to perform ultra-short-term prediction of power load based on the obtained optimal historical daily load curve and a preset prediction model; Among them, the preset prediction model adopts a combined prediction model, constructs an ultra-short-term prediction model through the extreme learning machine, and combines the Bayesian optimization extreme learning machine to achieve ultra-short-term prediction of power load; The power load sequence in the optimal historical daily load curve is decomposed into power load cycle components, power load trend components, and power load residual components through the time series decomposition method to obtain fine-grained data. The spatiotemporal characteristics of the power load trend components are learned using convolutional neural networks and long short-term memory neural networks to obtain the load trend forecast value for the next day. In the process of obtaining the load trend forecast value for the next day, the power load sequence in the optimal historical daily load curve is trained through a hybrid deep learning model, and the output value of its output layer is the load trend forecast value for the next day; the hybrid deep learning model includes a time series decomposition module, a convolutional neural network spatial feature extraction module, a long short-term memory neural network time feature extraction module and an output layer.

7. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the load ultra-short-term forecasting method based on two-stage intelligent feature engineering as described in any one of claims 1 to 5 are implemented.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps in the ultra-short-term load forecasting method based on two-stage intelligent feature engineering are implemented as described in any one of claims 1 to 5.

Images