Day-ahead time-sharing electric quantity prediction method and device based on feature fusion and medium
By constructing a three-dimensional power matrix and a pyramid ConvLSTM model, the problems of insufficient accuracy and generalization ability of traditional power forecasting methods are solved, and accurate prediction of day-ahead time-of-use power consumption is achieved, thereby improving the operating efficiency and reliability of the power system.
Patent Information
- Application Number
- CN202411187042.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-11-22
AI Technical Summary
Traditional power generation forecasting methods are unable to meet the diverse and high-precision forecasting needs. They suffer from problems such as missing power generation data, small data samples, and increased model complexity leading to reduced interpretability. Furthermore, some models have insufficient generalization ability on new datasets.
A three-dimensional power matrix is constructed, and dominant features are selected by combining machine learning and evidence optimization algorithms. A pyramid convolutional long short-term memory network (ConvLSTM) model is used for multi-scale feature extraction and temporal feature capture to achieve power prediction.
It improves the accuracy of electricity forecasting and the generalization ability of models, reduces electricity costs, promotes the operating efficiency and reliability of the power system, and adapts to real-time changes in the electricity market.
Smart Images

Figure CN119004035A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of electricity forecasting, and in particular to a method, device and medium for day-ahead time-sharing electricity forecasting based on feature fusion. Background Art
[0002] With the development of society and economic growth, energy consumption has become a global problem, and electricity demand is becoming more and more complex. Traditional electricity forecasting methods often fail to meet the diverse and high-precision forecasting needs. In the context of power market reform, electricity forecasting has become a key technology. By accurately forecasting electricity demand, it is possible to optimize resource allocation, plan supply, and improve the rationality of power generation plans. The main goal of day-ahead electricity forecasting based on multi-dimensional and multi-scale feature fusion is to predict electricity data in order to achieve accurate electricity forecasting in a specific time period in the future. This method integrates multi-dimensional data features and different time scales to improve the accuracy of day-ahead electricity forecasting.
[0003] The time-based electricity forecasting method based on advanced technology has become one of the hot topics in current research. The time-based electricity forecasting technology has developed rapidly. The use of machine learning and deep learning can effectively capture the complex relationships in the data. Commonly used time-based electricity forecasting models include autoregressive integrated moving average model (ARIMA), support vector machine (SVM), long short-term memory network (LSTM) and convolutional neural network (CNN). In order to improve the stability of the forecast, multiple different forecasting models are also integrated into an integrated model or the optimal parameters are found for a single model. Conventional machine learning methods such as SVM model parameter tuning are complex and have poor performance in processing large-scale data; using deep learning methods such as LSTM models, the amount of training data required is huge and the model interpretability is poor. Therefore, in the continuous development of day-ahead time-segment electricity forecasting, there are still some problems, such as: missing electricity data, small number of data samples, etc.; with the introduction of deep learning, the complexity of the prediction model increases, resulting in reduced interpretability of the model; some models only perform well for some data sets, and lack generalization capabilities on new data sets, affecting the accuracy of electricity forecasting. Summary of the invention
[0004] The purpose of this application is to provide a method, device and medium for predicting time-sharing electricity consumption before the day is reached based on feature fusion, which can realize accurate prediction of time-sharing electricity consumption before the day is reached.
[0005] To achieve the above objectives, this application provides the following solutions:
[0006] In a first aspect, the present application provides a method for predicting day-ahead time-of-day electricity consumption based on feature fusion, including:
[0007] A three-dimensional electricity matrix is constructed based on historical electricity data; the three-dimensional electricity matrix includes a number of single-day time-sharing electricity feature matrices merged and spliced in time order, each single-day time-sharing electricity feature matrix includes characteristic variables that affect electricity prediction under all time-sharing electricity periods of a single day; the characteristic variables that affect electricity prediction include time characteristics, meteorological characteristics, user behavior characteristics and historical electricity characteristics;
[0008] Based on machine learning and evidence optimization combination algorithm, the characteristic variables that affect power forecast are screened to obtain the dominant feature set;
[0009] According to the dominant feature set, the three-dimensional power matrix is reduced in dimension to obtain the three-dimensional matrix of time-sharing power;
[0010] Constructing a time-sharing electricity prediction model based on a pyramid convolutional long short-term memory network (Convolutional LSTM Network, ConvLSTM); the time-sharing electricity prediction model based on the pyramid ConvLSTM includes: a time-sharing electricity multi-scale feature extraction module based on a pyramid structure, an electricity time series feature capture module based on ConvLSTM, and an electricity prediction module based on fusion features;
[0011] The pyramid ConvLSTM-based time-sharing electricity prediction model is used to predict the time-sharing electricity value of the next day according to the time-sharing electricity three-dimensional matrix.
[0012] Optionally, a three-dimensional power matrix is constructed based on historical power data, specifically including:
[0013] Acquire historical electricity data; the historical electricity data includes a number of single-day electricity consumption sequences within a historical time period and corresponding characteristic variable data affecting electricity prediction;
[0014] The daily electricity consumption sequence is divided into multiple subsequences along the time axis according to the time-sharing electricity price table, and the cumulative electricity of each subsequence is calculated in turn to obtain the daily time-sharing electricity information;
[0015] According to the daily time-of-use electricity information and the corresponding characteristic variable data that affect electricity prediction, all discrete variables are processed into independent and mutually exclusive features using one-hot encoding, and all continuous variables are mapped to the [0,1] interval using the deviation standardization algorithm to obtain the daily time-of-use electricity feature matrix;
[0016] All single-day time-sharing electricity characteristic matrices in the historical time period are merged and spliced in chronological order to obtain a three-dimensional electricity matrix.
[0017] Optionally, based on machine learning and evidence optimization combination algorithm, characteristic variables affecting power forecast are screened to obtain a dominant feature set, specifically including:
[0018] The filtering method, embedding method and wrapping method in machine learning are used to calculate the characteristic contribution of characteristic variables affecting power prediction respectively, and the characteristic contribution of each characteristic variable in a single evaluation system is obtained;
[0019] The DS evidence theory is introduced to combine the characteristic contribution of each characteristic variable in a single evaluation system, and the characteristic contribution of each characteristic variable in the set of all evaluation systems is obtained;
[0020] The feature contribution of each feature variable in all evaluation system sets is sorted from large to small, and the feature cumulative contribution rate is calculated to screen the dominant features and obtain the dominant feature set.
[0021] Optionally, the filtering method is a Pearson correlation coefficient algorithm; the embedding method is an extreme gradient boosting (EXtreme Gradient Boosting, XGBoost) algorithm; the wrapping method is a deep neural network (Deep Neural Network, DNN) algorithm.
[0022] Optionally, the pyramid-structured time-sharing electricity multi-scale feature extraction module is composed of a cascade of convolutional networks with a convolution kernel of 1×1 and three convolution kernels of 3×3, and each convolutional network corresponds to an output layer connected to a downstream network; the pyramid-structured time-sharing electricity multi-scale feature extraction module is used to receive a three-dimensional matrix of time-sharing electricity and output four time-sharing electricity feature matrices of different scales.
[0023] Optionally, the ConvLSTM-based electricity timing feature capture module is composed of four independent encoding channels, each encoding channel is cascaded by a layer of ConvLSTM network and a layer of convolutional network; the ConvLSTM-based electricity timing feature capture module is used to receive four time-sharing electricity feature matrices of different scales and output corresponding timing change characteristics.
[0024] Optionally, the electricity consumption prediction module based on fusion features includes a feature fusion layer and a fully connected layer; the feature fusion layer is used to splice all time series change features along the channel dimension and merge the channels to obtain multi-scale fusion features, and the fully connected layer is used to receive the multi-scale fusion features and output the time-sharing electricity consumption value for the next day.
[0025] Optionally, the time characteristics include: holiday information, season information, month information, week information of the electricity to be predicted and the total duration of the time-sharing electricity period corresponding to the electricity to be predicted; the meteorological characteristics include: the temperature, air pressure, precipitation, irradiance, wind speed and wind direction in the user's area; the user behavior characteristics include: the total duration of the user's main activities in the time-sharing electricity period corresponding to the electricity to be predicted; the historical electricity characteristics include: electricity data of the week before the electricity to be predicted.
[0026] In a second aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the feature fusion-based day-ahead time-of-day electricity forecasting method.
[0027] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting day-ahead time-of-day electricity consumption based on feature fusion.
[0028] According to the specific embodiments provided in this application, this application discloses the following technical effects:
[0029] The present application provides a method, device and medium for predicting day-ahead time-of-day electricity based on feature fusion. First, a three-dimensional electricity matrix is constructed based on historical electricity data, and the natural factors and social behavior characteristics affecting electricity are comprehensively analyzed, thereby broadening the characteristic dimension of electricity. Secondly, based on machine learning and evidence optimization combination algorithms, characteristic variables affecting electricity prediction are screened to obtain a dominant feature set, and the dominant features affecting electricity in each period are comprehensively screened from multiple evaluation levels, so that the formed multi-dimensional feature set of time-of-day electricity is more accurate, which can improve the prediction efficiency of subsequent models. In addition, a time-of-day electricity prediction model based on pyramid ConvLSTM is constructed. On the one hand, a pyramid-structured time-of-day electricity multi-scale feature extraction technology is used to extract multiple different-scale electricity features, which can not only more comprehensively perceive the correlation between the dominant features and electricity, but also fully learn the coupling relationship between electricity in different periods. On the other hand, the ConvLSTM-based electricity time series feature capture technology can simultaneously extract multi-scale electricity features with day granularity and time period granularity, obtain multi-dimensional and multi-scale fused time series features, and improve the feature learning ability of the prediction model. Therefore, this application can achieve accurate prediction of day-ahead time-of-day electricity consumption, improve the operating efficiency and reliability of the power system, reduce electricity costs, and promote sustainable development to cope with real-time changes in the power market. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0031] Figure 1 A flowchart of the day-ahead time-sharing electricity quantity prediction method based on feature fusion provided in this application;
[0032] Figure 2 A schematic diagram of the daily electricity consumption division provided for this application;
[0033] Figure 3 A schematic diagram of the three-dimensional power matrix provided for this application;
[0034] Figure 4 The dominant feature recognition graph based on the combined algorithm of machine learning and evidence optimization provided for this application;
[0035] Figure 5 The structural diagram of the time-sharing power prediction model based on pyramid ConvLSTM provided in this application;
[0036] Figure 6 This is a technical architecture diagram of the day-ahead time-of-day electricity forecasting method based on feature fusion provided in this application. DETAILED DESCRIPTION
[0037] 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, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0038] In order to solve the problems existing in the prior art, this application proposes a method, device and medium for predicting time-sharing electricity consumption based on feature fusion. This method comprehensively analyzes the natural factors and social behavior characteristics that affect electricity consumption, constructs a three-dimensional matrix of time-sharing electricity consumption, and realizes multi-dimensional analysis of electricity fluctuations. It uses a dominant feature identification method based on machine learning and evidence optimization combination algorithm to accurately identify the dominant features that affect each time-sharing electricity period from multiple dimensions. After the features are input into the pyramid convolution network to extract multi-scale features, the inherent time series features are mined based on the ConvLSTM network. The electricity consumption prediction value for the time period before the day is obtained through the electricity prediction technology based on fusion features, which can realize accurate prediction of the time-sharing electricity consumption before the day.
[0039] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.
[0040] like Figure 1 As shown, the day-ahead time-sharing electricity consumption prediction method based on feature fusion provided in the present application includes the following steps 1 to 5.
[0041] Step 1: Construct a three-dimensional power matrix based on historical power data. The three-dimensional power matrix includes several single-day time-sharing power feature matrices merged and spliced in time order, and each single-day time-sharing power feature matrix includes characteristic variables that affect power forecasting under all time-sharing power periods of a single day. The characteristic variables that affect power forecasting include time characteristics, meteorological characteristics, user behavior characteristics, and historical power characteristics.
[0042] This step specifically includes: obtaining historical electricity data, the historical electricity data including several single-day electricity consumption sequences within the historical time period and corresponding characteristic variable data affecting electricity prediction; dividing the single-day electricity consumption sequence into multiple subsequences along the time axis based on the time-sharing electricity price table, calculating the cumulative electricity of each subsequence in turn, and obtaining the single-day time-sharing electricity information. The division process is as follows: Figure 2 As shown in the figure; according to the single-day time-sharing electricity information and the corresponding characteristic variable data affecting electricity prediction, all discrete variables are processed into independent and mutually exclusive features using the unique hot encoding method, and all continuous variables are mapped to the [0,1] interval using the deviation standardization algorithm to obtain the single-day time-sharing electricity feature matrix; all single-day time-sharing electricity feature matrices in the historical time period are merged and spliced in chronological order to obtain a three-dimensional electricity matrix, as shown in the figure. Figure 3 shown.
[0043] Exemplarily, this application summarizes the characteristic variables that affect power forecasting into four categories: time characteristics, meteorological characteristics, user behavior characteristics, and historical power characteristics based on existing theoretical research and exploratory experimental experience. Time characteristics include holiday information, season information, month information, week information, and the total duration of the corresponding period of the power to be predicted. Meteorological characteristics include basic elements such as temperature, air pressure, precipitation, irradiance, wind speed, and wind direction in the user's area, among which the precipitation feature is the cumulative precipitation in the corresponding period, and the remaining features are the average values of the corresponding period of the feature. User behavior characteristics are the total duration of the user's main activities in the corresponding period of the power to be predicted (such as the home time of residential users, the production time of factory users, etc.). Historical power characteristics are the power data of the previous week of the power to be predicted. All the above discrete variables are processed into independent and mutually exclusive features using the unique hot encoding method, and all continuous variables are mapped to the [0,1] interval using the deviation standardization algorithm to obtain a single-day time-sharing power feature matrix with the time-sharing power period as the X-axis and the feature as the Y-axis, as shown in formula (1).
[0044]
[0045] In the formula, x 11 Represents the time-sharing electricity period is x 1 , with the feature y 1 The data points are as follows, n and w represent the number of time-sharing electricity periods and the number of characteristics respectively.
[0046] Step 2: Based on machine learning and evidence optimization combination algorithm, the characteristic variables that affect power forecast are screened to obtain the dominant feature set.
[0047] This step specifically includes: using the filtering method, embedding method and wrapping method in machine learning to respectively calculate the feature contribution of the feature variables that affect the electricity prediction, and obtain the feature contribution of each feature variable in a single evaluation system; introducing the DS evidence theory to combine the feature contribution of each feature variable in a single evaluation system, and obtain the feature contribution of each feature variable in the set of all evaluation systems; sorting the feature contribution of each feature variable in the set of all evaluation systems from large to small, calculating the feature cumulative contribution rate to screen the dominant features, and obtaining the dominant feature set.
[0048] For example, the present application proposes a feature screening method of filtering, embedding and wrapping, selects Pearson correlation coefficient algorithm, XGBoost algorithm and DNN algorithm to evaluate the contribution of each feature to each type of electricity in turn, then combines each evaluation system through DS evidence theory, and finally identifies the dominant feature by judging the cumulative contribution rate. The process is as follows: Figure 4 shown.
[0049] Firstly, the Pearson correlation coefficient algorithm is used to evaluate the waveform similarity index between each characteristic variable and the electricity quantity in a certain period, and then the contribution of each characteristic to the electricity quantity prediction is quantified. The calculation process is shown in formulas (2)-(3).
[0050]
[0051] In the formula, α i is the i-th feature variable X i Pearson correlation coefficient with the electric quantity Y, and Represents the characteristic variable X i and the mean value of the electric quantity Y, is the pth number in the ith characteristic variable, Y p is the pth number in the electric quantity sequence. i The value range of is [-1,1]. The closer its absolute value is to 1, the higher the correlation between this type of feature and the electricity consumption in this period. is the contribution of the i-th characteristic variable to the power forecast for this period under the Pearson correlation coefficient algorithm, N is the number of characteristics, and
[0052] Secondly, the XGBoost model’s built-in feature contribution index is used to quantify the importance of each feature variable to the amount of electricity during this period. The XGBoost model calculates the contribution of each feature variable based on the built-in feature importance function. After the model training is completed, it is directly output, and there is
[0053] Subsequently, all characteristic variables of the electricity consumption in this period are combined and predicted through the DNN model. The electricity consumption prediction accuracy is used as the model performance evaluation standard, and the characteristic Shapley value is calculated to quantify the feature contribution. The calculation process is shown in Equations (4)-(7).
[0054]
[0055] V=1-E (5)
[0056]
[0057] Where E is the prediction error of the DNN model, M is the total number of samples, is the true value of the power of the jth sample, is the power prediction value of the jth sample, V is the power prediction accuracy; η i is the i-th feature variable X i The Shapley value of O is the feature set of the input model, {X 1 ,…,X i-1 ,X i+1 ,…,X N} represents that the feature set does not contain feature variable X i , |O| is the number of features in the feature subset; is the contribution of the i-th characteristic variable to the power prediction under the DNN algorithm,
[0058] Afterwards, the DS evidence theory is introduced to combine the feature contributions in the above three evaluation systems to obtain the final feature contribution. The calculation process is shown in equations (8)-(10).
[0059]
[0060] In the formula, A is the evaluation system set; A r is the focal element, corresponding to a single evaluation index; θ r (A r) and θ(A) represent the feature contribution in a single evaluation system and in the set of all evaluation systems respectively; K is the conflict coefficient, which is positively correlated with the difference in feature contribution in different systems.
[0061] Finally, all features are sorted from large to small according to the final feature contribution, and the cumulative contribution rate of the first m dominant features is calculated. The cumulative contribution rate increment of the feature is determined by comparing it with the preset increment threshold Δρ. 0 The size relationship of is determined by screening out the top m dominant features to determine the final feature set of the electricity consumption in this period. The calculation process is shown in formulas (11)-(13).
[0062]
[0063] Δρ m =ρ m -ρ m-1 (12)
[0064] Δρ m ≤Δρ 0 ≤Δρ m-1 (13)
[0065] In the formula, ρ m Indicates the contribution ratio of the first m dominant features in all feature sets, θ i Indicates the contribution of each feature.
[0066] After screening the features of the electricity in each time period in turn according to the above method, the final feature combination, i.e., the dominant feature set, is obtained by taking the union of the feature sets of the electricity in all time periods.
[0067] Step 3: Reduce the dimension of the three-dimensional power matrix according to the dominant feature set to obtain the time-sharing power three-dimensional matrix.
[0068] Specifically, along the characteristic dimension of the three-dimensional electricity matrix, characteristic variables other than the dominant characteristic set are eliminated to obtain the three-dimensional matrix of time-sharing electricity.
[0069] Step 4: Build a time-sharing power consumption prediction model based on pyramid ConvLSTM. Figure 5 As shown, the pyramid ConvLSTM-based time-sharing power forecasting model includes: a pyramid-structured time-sharing power multi-scale feature extraction module, a ConvLSTM-based power time series feature capture module, and a fusion feature-based power forecasting module. The above modules are introduced in detail below.
[0070] 1) Multi-scale feature extraction module of time-sharing electricity based on pyramid structure
[0071] The pyramid-structured multi-scale feature extraction module for time-sharing electricity is composed of a 1×1 convolution kernel and three 3×3 convolution kernels cascaded together. Each convolution network corresponds to an output layer connected to the downstream network. The module divides the input three-dimensional matrix of time-sharing electricity into L two-dimensional feature matrices according to the time axis, where L is the time scale of the input data. Then, the convolution operation is performed on the electricity at each moment in turn to obtain hidden three-dimensional features with a fragment length of L. As the convolution network deepens, four time-sharing electricity feature matrices of different scales are finally output.
[0072] 2) Power time series feature capture module based on ConvLSTM
[0073] The ConvLSTM-based power time series feature capture module consists of four independent encoding channels, each of which is composed of a layer of ConvLSTM network and a layer of convolutional network cascaded. One encoding channel receives a scaled time-sharing power feature matrix output by the above feature extraction module, and outputs the time series change features of the scaled time-sharing power feature matrix.
[0074] The single channel encoding process is as follows: First, the incoming time-sharing power feature matrix is passed into the ConvLSTM network to learn the time series development characteristics of power. Similar to the LSTM model, the ConvLSTM model controls the input, forgetting, updating and output of data through the gating unit. The input of each moment of this model is a two-dimensional matrix, and the two-dimensional information of the future moment is obtained by inferring the hidden state of the current moment input and the adjacent data of the past moment. The specific process is shown in equations (14)-(18).
[0075]
[0076] In the formula, is the convolution kernel weight, * represents convolution, represents element-wise multiplication, x t is the input two-dimensional matrix, h t-1 is the hidden state two-dimensional matrix at time t-1 (i.e. the previous moment), c t-1 is the two-dimensional matrix of memory state at time t-1, I t is the input gate activation value at time t, T xi , T hi , l i is the weight and bias of the input gate, T ci is the weight matrix from the memory state at time t-1 to the input gate, F t is the forget gate activation value at time t, T xf , T hf , l f is the forget gate weight and bias term, T cfis the weight matrix from the memory state at time t-1 to the forget gate, C t is the cell state at time t, f t is the forget gate unit at time t, i t is the input gate unit at time t, T xc , T hc , l c To update the gate weights and bias terms, O t is the output gate activation value at time t, T xo , T ho , l o is the weight and bias term of the output gate, T co is the weight matrix from the memory state at time t (i.e. the current time) to the output gate, H t is the hidden state at time t, c t is the two-dimensional matrix of memory state at time t, o t is the output gate unit at time t.
[0077] Subsequently, the ConvLSTM layer uses the learned L-1 hidden states as the power change features learned at the Lth moment, and convolves them along the power dimension through a 1×q convolution kernel to obtain a 1×w two-dimensional power hidden feature, where q is the number of power features and w is the number of time periods for power.
[0078] 3) Power prediction module based on fusion features
[0079] The electricity prediction module based on fusion features includes a feature fusion layer and a fully connected layer. First, the feature fusion layer splices multiple two-dimensional features obtained by the time series feature capture module along the channel dimension, and then uses a 1×1 convolution kernel to merge the channels to obtain multi-scale fusion features. Finally, the fully connected layer obtains the electricity values of different time periods in the future day based on the multi-scale fusion features. In addition, according to the electricity price information corresponding to the time period, the electricity of the same electricity price can be accumulated and summed to obtain the electricity forecast value of the future day's time-sharing electricity price.
[0080] Step 5: Use the pyramid ConvLSTM-based time-sharing electricity prediction model to predict the time-sharing electricity value for the next day according to the time-sharing electricity three-dimensional matrix.
[0081] like Figure 6As shown, the day-ahead time-of-use electricity forecasting method based on feature fusion proposed in this application involves a day-ahead time-of-use electricity forecasting technology based on multi-dimensional and multi-scale feature fusion, which is composed of a dominant feature identification method of fusion filtering method, wrapping method and embedding method and a time-of-use electricity forecasting technology based on pyramid ConvLSTM. The dominant feature identification method of fusion filtering method, wrapping method and embedding method includes two parts: a three-dimensional electricity matrix construction method based on time-of-use electricity price and a dominant feature identification method based on machine learning and evidence optimization combination algorithm. The time-of-use electricity forecasting technology based on pyramid ConvLSTM includes three parts: a time-of-use electricity multi-scale feature extraction technology based on pyramid structure, a power time series feature capture technology based on ConvLSTM, and an electricity forecasting technology based on fusion features. This application first constructs a multi-period electricity feature matrix by the "three-dimensional electricity matrix construction method based on time-of-use electricity price", and then screens out the dominant features of electricity in each period according to the "dominant feature identification method based on machine learning and evidence optimization combination algorithm", and then extracts electricity features of different scales through the "time-of-use electricity multi-scale feature extraction technology based on pyramid structure", and then uses the "electricity time series feature capture technology based on ConvLSTM" to mine the hidden time series features of multi-scale electricity features, and finally all multi-scale electricity time series features are obtained through the "electricity prediction technology based on fusion features" to obtain the predicted value of time-of-use electricity a day ago.
[0082] Compared with the existing technology, this application expands the latitude and information content of training data through multi-dimensional feature fusion and multi-scale feature extraction, enriching the diversity of data under the premise of a small number of samples; this application is based on the dominant feature recognition method of machine learning and evidence optimization combination algorithm, which can clarify the contribution of different features to improve the interpretability of the model; this application uses time-sharing electricity multi-scale feature extraction technology and electricity time series feature capture technology to improve the robustness of the model, so that the model can better adapt to different data patterns and changes, and improve generalization ability.
[0083] In summary, in view of the problems and defects of the model's time-division electricity forecast, this application uses the time series data and corresponding feature information in the power system to construct a three-dimensional electricity matrix to better reflect the energy consumption of the power system. By constructing a three-dimensional electricity matrix, the energy consumption of the power system can be analyzed and understood more comprehensively, providing an important reference for the dispatch, operation and management of the power system; in the machine learning and evidence optimization algorithm, machine learning and evidence optimization are combined to provide more reliable decisions and predictions for the model. At the same time, the combined algorithm can overcome problems such as scarcity of data samples and improve the model's adaptability and overall generalization ability to input data. The time-division electricity time series feature extraction technology based on the feature graph pyramid is one of the key technologies in the time-division electricity forecast. Its structure allows multi-scale and multi-level feature extraction of electricity data, which can effectively capture the changing laws of data in the time dimension. Through this method, the periodicity, trend and abnormal changes of electricity data can be analyzed more carefully, providing an important data basis for subsequent predictions. The ConvLSTM-based electricity time series feature capture technology utilizes the ConvLSTM model, which integrates the CNN and LSTM structures. CNN is used to extract spatial features from spatiotemporal data, while LSTM is used to perform time series modeling on the extracted features. It introduces convolution operations in the recurrent hidden layer of LSTM, allowing the network to process spatiotemporal information simultaneously, thereby better capturing the spatiotemporal correlation in time series data. This application can achieve accurate prediction of day-ahead time-based electricity consumption, improve the operating efficiency and reliability of the power system, reduce electricity costs, and promote sustainable development to cope with real-time changes in the power market. With the continuous optimization and development of data collection, processing technology, and prediction models, the time-based electricity forecasting method will continue to promote technological innovation and development in the power field, and help promote the energy system to develop in a more intelligent, efficient, and reliable direction.
[0084] In an exemplary embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.
[0085] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.
[0086] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.
[0087] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.
[0088] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).
[0089] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.
[0090] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0091] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.
Claims
1. A method for predicting time-based electricity consumption based on feature fusion, characterized in that: include: A three-dimensional electricity matrix is constructed based on historical electricity data; the three-dimensional electricity matrix includes a number of single-day time-sharing electricity feature matrices merged and spliced in time order, each single-day time-sharing electricity feature matrix includes characteristic variables that affect electricity prediction under all time-sharing electricity periods of a single day; the characteristic variables that affect electricity prediction include time characteristics, meteorological characteristics, user behavior characteristics and historical electricity characteristics; Based on machine learning and evidence optimization combination algorithm, the characteristic variables that affect power forecast are screened to obtain the dominant feature set; According to the dominant feature set, the three-dimensional power matrix is reduced in dimension to obtain the three-dimensional matrix of time-sharing power; Constructing a time-sharing electricity prediction model based on a pyramid convolutional long short-term memory network; the time-sharing electricity prediction model based on a pyramid convolutional long short-term memory network includes: a time-sharing electricity multi-scale feature extraction module based on a pyramid structure, an electricity time series feature capture module based on a convolutional long short-term memory network, and an electricity prediction module based on fusion features; A time-sharing electricity prediction model based on pyramid convolutional long short-term memory network is used to predict the time-sharing electricity value of the next day according to the time-sharing electricity three-dimensional matrix.
2. The method for predicting power consumption based on time-sharing and feature fusion before the day according to claim 1 is characterized in that: Construct a three-dimensional power matrix based on historical power data, including: Acquire historical electricity data; the historical electricity data includes a number of single-day electricity consumption sequences within a historical time period and corresponding characteristic variable data affecting electricity prediction; The daily electricity consumption sequence is divided into multiple subsequences along the time axis according to the time-sharing electricity price table, and the cumulative electricity of each subsequence is calculated in turn to obtain the daily time-sharing electricity information; According to the daily time-of-use electricity information and the corresponding characteristic variable data that affect electricity prediction, all discrete variables are processed into independent and mutually exclusive features using one-hot encoding, and all continuous variables are mapped to the [0,1] interval using the deviation standardization algorithm to obtain the daily time-of-use electricity feature matrix; All single-day time-sharing electricity characteristic matrices in the historical time period are merged and spliced in chronological order to obtain a three-dimensional electricity matrix.
3. The method for predicting power consumption based on time-sharing and feature fusion before the day according to claim 1 is characterized in that: Based on machine learning and evidence optimization combination algorithm, the characteristic variables that affect power forecast are screened to obtain the dominant feature set, including: The filtering method, embedding method and wrapping method in machine learning are used to calculate the characteristic contribution of characteristic variables affecting power prediction respectively, and the characteristic contribution of each characteristic variable in a single evaluation system is obtained; The DS evidence theory is introduced to combine the characteristic contribution of each characteristic variable in a single evaluation system, and the characteristic contribution of each characteristic variable in the set of all evaluation systems is obtained; The feature contribution of each feature variable in all evaluation system sets is sorted from large to small, and the feature cumulative contribution rate is calculated to screen the dominant features and obtain the dominant feature set.
4. The method for predicting power consumption based on time-sharing based on feature fusion according to claim 3 is characterized in that: The filtering method is the Pearson correlation coefficient algorithm; the embedding method is the extreme gradient boosting algorithm; and the wrapping method is the deep neural network algorithm.
5. The method for predicting power consumption based on time-sharing and feature fusion before the day according to claim 1 is characterized in that: The pyramid-structured time-sharing electricity multi-scale feature extraction module is composed of a cascade of convolutional networks with a convolution kernel of 1×1 and three convolutional kernels of 3×3, and each convolutional network corresponds to an output layer connected to a downstream network; the pyramid-structured time-sharing electricity multi-scale feature extraction module is used to receive a three-dimensional matrix of time-sharing electricity and output four time-sharing electricity feature matrices of different scales.
6. The method for predicting power consumption based on time-sharing and feature fusion before the day according to claim 5 is characterized in that: The electricity time series feature capture module based on convolutional long short-term memory network is composed of four independent encoding channels, each encoding channel is cascaded by a layer of convolutional long short-term memory network and a layer of convolutional network; the electricity time series feature capture module based on convolutional long short-term memory network is used to receive four time-sharing electricity feature matrices of different scales and output corresponding time series change characteristics.
7. The method for predicting power consumption based on time-sharing and feature fusion according to claim 6 is characterized in that: The power prediction module based on fusion features includes a feature fusion layer and a fully connected layer; The feature fusion layer is used to splice all time series change features along the channel dimension and merge the channels to obtain multi-scale fusion features. The fully connected layer is used to receive the multi-scale fusion features and output the time-sharing power value for the next day.
8. The method for predicting power consumption based on time-sharing and feature fusion before the day according to claim 1 is characterized in that: The time characteristics include: holiday information, season information, month information, week information of the electricity to be predicted and the total duration of the time-sharing electricity period corresponding to the electricity to be predicted; the meteorological characteristics include: the temperature, air pressure, precipitation, irradiance, wind speed and wind direction in the user's area; the user behavior characteristics include: the total duration of the user's main activities in the time-sharing electricity period corresponding to the electricity to be predicted; the historical electricity characteristics include: electricity data of the week before the electricity to be predicted.
9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the day-ahead time-of-day electricity forecasting method based on feature fusion as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting day-ahead time-of-day electricity consumption based on feature fusion described in any one of claims 1 to 8 is implemented.
Citation Information
Cited By
Time-sharing electric quantity prediction method
CN119940959A
