CN115689026A - Method and system for short-term prediction of house load - Google Patents
Method and system for short-term prediction of house load Download PDFInfo
- Publication number
- CN115689026A CN115689026A CN202211377095.XA CN202211377095A CN115689026A CN 115689026 A CN115689026 A CN 115689026A CN 202211377095 A CN202211377095 A CN 202211377095A CN 115689026 A CN115689026 A CN 115689026A
- Authority
- CN
- China
- Prior art keywords
- load
- data
- prediction
- attention mechanism
- load curve
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a method and a system for predicting a house load in a short term, and belongs to the technical field of power grid dispatching. Aiming at the problem of large difference between the residential electricity peak value and the average value, the load value to be predicted is decomposed into a basic load curve and a difference value; the determination of the base load curve is mainly considered from two aspects: one is the periodicity of the daily load curve of the house and the other is the validity of the load curve. And a reference load curve is constructed, so that the difference of load peak values and mean values is effectively reduced, and the prediction accuracy is improved. Meanwhile, the self-attention mechanism and the external attention mechanism are combined, and after big data training, the model can more effectively extract time sequence characteristics.
Description
Technical Field
The invention belongs to the technical field of power grid dispatching, and particularly relates to a method and a system for short-term prediction of residential load.
Background
In global energy consumption, the proportion of electricity used by buildings has reached 40%, and the electricity used by household residences occupies a large part of the electricity. Because the power utilization mode of household power utilization is changeable, the household power utilization mode is uncertain in the peak power utilization period of a day, and certain difficulty can be brought to peak clipping and valley filling of a power grid. However, if the load of the household residence can be accurately predicted, the power grid can make some incentive demand response plans for users in peak periods, so that the users actively avoid the own power utilization periods and the power utilization peak periods, the peak clipping and valley filling effects can be achieved, the stability of the power grid is improved, and the economic operation of the power grid is guaranteed.
The load of the family residence is influenced by various factors, such as external weather, festivals and holidays, the conditions are complex and changeable, the fluctuation of a load curve is large, the nonlinear characteristic is obvious, and great difficulty is brought to prediction. Moreover, the peak value and the average value of the residential electricity are greatly different, so that the accuracy of the peak value prediction by a general prediction model is poor. For the problem of large fluctuation of the house load, the fluctuation can be reduced by means of load decomposition or normalization. The load decomposition is to decompose the load sequence into a plurality of subsequences with regular signals, and the sum of the subsequences can restore the initial sequence. The number of subsequences may be determined based on the size of the volatility. For the load with large fluctuation, the influence of high-frequency noise can be effectively relieved by decomposing the load into a plurality of subsequences. However, the decomposed sequence cannot completely reduce the original sequence, a certain error exists after reduction, the processing time is positively correlated with the number of subsequences, so that the decomposition cannot be carried out infinitely, and some subsequences still have the problem of large peak value and mean value difference. All load sequences are directly normalized and then predicted, the proportion of the peak value and the average value is unchanged, the problem of large difference between the peak value and the average value is not solved, and the prediction effect on the peak value after reverse normalization is still poor.
Disclosure of Invention
In view of the above defects or improvement needs of the prior art, the present invention provides a method and a system for short-term prediction of residential load, which aim to solve the technical problem of poor prediction accuracy caused by large difference between peak value and mean value of residential electricity consumption.
To achieve the above object, according to an aspect of the present invention, there is provided a residential load short-term prediction method, including:
s1, performing variational modal decomposition on historical load data subjected to data cleaning to serve as a training set;
s2, performing primary segmentation on the historical load data subjected to data cleaning according to natural weeks, averaging all subsequences subjected to primary segmentation, and performing secondary segmentation on the average sequence by taking days as a unit;
s4, taking the average value sequence as a typical load curve of each day, selecting load data N days before the date to be predicted, and calculating by taking the similarity between the load data N days before and the typical load curve as a weight to obtain a reference load curve of the date to be predicted;
s5, sequentially fetching data from the training set by adopting a sliding window method and inputting the data into a prediction network for iterative training to obtain a trained prediction model; the prediction network comprises a self-attention mechanism, an external attention mechanism and a time convolution network;
the self-attention mechanism is used for distributing weights to the input data according to the internal features of the input data to enable the internal key feature weight to be increased and the weights of other features to be decreased so as to help the prediction model to extract key features;
an external attention mechanism, which is used for distributing weight to the current input data according to the characteristics of all the input data in the training set, so that the periodically related characteristic weight is increased, and the other characteristic weights are decreased;
the time convolution network is used for learning the data processed by the self-attention mechanism and the external attention mechanism, changing the value of an internal parameter through error back propagation, extracting the periodicity and the time dependency of the data and outputting a prediction result;
and S6, inputting the historical load data of N days before the measured day into the trained prediction model to obtain a load prediction result, and superposing the load prediction result and a reference load curve to obtain a load prediction value of the next day.
Further, the data cleaning comprises filling null values in the data and removing the data with the load less than a certain threshold value within a set time length.
Further, the reference load curve is calculated by the following formula:
The invention also provides a short-term prediction system for residential load, which comprises:
the first data preprocessing module is used for performing variation modal decomposition on the historical load data subjected to data cleaning to serve as a training set;
the second data preprocessing module is used for carrying out primary segmentation on the historical load data subjected to data cleaning according to natural weeks, solving the average value of all subsequences of the primary segmentation, and carrying out secondary segmentation on the average value sequence by taking days as a unit;
the reference coincidence curve building module is used for taking the average value sequence as a typical load curve of each day, selecting load data N days before the date to be predicted, and calculating the similarity between the load data N days before and the typical load curve as a weight to obtain a reference load curve of the date to be predicted;
the prediction model training module is used for sequentially taking data from a training set by adopting a sliding window method and inputting the data into a prediction network for iterative training to obtain a trained prediction model; the prediction network comprises a self-attention mechanism, an external attention mechanism and a time convolution network;
the self-attention mechanism is used for distributing weights to input data according to internal features of the input data to enable the internal key feature weight to be increased and the weights of other features to be reduced, and helps the prediction model to extract key features;
the external attention mechanism is used for distributing weight to the current input data according to the characteristics of all input data in the training set, so that the periodically related characteristic weight is increased, and the other characteristic weights are reduced;
the time convolution network is used for learning the data processed by the self-attention mechanism and the external attention mechanism, changing the value of an internal parameter through error back propagation, extracting the periodicity and the time dependency of the data and outputting a prediction result;
and the short-term load prediction module is used for inputting the historical load data of N days before the measured day into the trained prediction model to obtain a load prediction result, and superposing the load prediction result and a reference load curve to obtain a load prediction value of the next day.
In general, the above technical solutions contemplated by the present invention can achieve the following advantageous effects compared to the prior art.
(1) According to the invention, by constructing the reference load curve, the prediction target is decomposed into the difference value between the actual load curve and the reference load curve, so that the difference between the load peak value and the average value is effectively reduced, and the prediction accuracy is improved.
(2) The invention combines the self-attention mechanism and the external attention mechanism, and after big data training, the model can more effectively extract time sequence characteristics.
Drawings
FIG. 1 is a framework for short term prediction of residential load;
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the respective embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
Aiming at the problem of large difference between the residential electricity peak value and the average value, the load value to be predicted is decomposed into a basic load curve and a difference value; the determination of the base load curve is mainly considered from two aspects: one is the periodicity of the daily load curve of the home and the other is the validity of the load curve. The electricity utilization patterns of residential users in a certain limited time period are similar, that is, the load curve of the residential during the time period shows a certain natural periodicity, usually a natural cycle, and after the load curve of the users in a long time is segmented according to the natural cycle, a load sequence set in a certain electricity utilization pattern can be obtained. Furthermore, the sequences in the set are segmented by taking a day as a period to obtain a smaller periodic sequence. According to the similarity degree of the sequences, a reference load curve can be predicted.
The validity of the load curve is to consider the data that the load is at a lower level for a long time, and the lower level for a long time indicates that the residential appliance is in a standby state in the time period and is not actually used. In this state, the load curve of the house does not have a value of 0, but interferes with the prediction, and this data needs to be removed.
(1) Data pre-processing
The data preprocessing comprises two parts: (i) Data cleaning and variational modal decomposition, which is used for providing input data for the prediction model; (ii) The data is cleaned and segmented with a week period, and then segmented again with a day period to obtain 7 daily load sequence subsets for calculating a reference load curve.
Data cleansing has two main roles. One is a null in the padding data. And filling null values in the data by adopting a linear interpolation mode to prevent the null values from influencing the calculation. And secondly, removing data with a load at a lower level for a long time in the data. The long-time low level indicates that all household electrical appliances are in a standby state in the time period and are not actually used. This portion of the data interferes with the results and needs to be removed.
The electricity usage behavior of residential users may exhibit a significant periodicity in units of weeks, which may be exploited by load divisions in cycles of weeks. The sequence of daily activities is not fixed during a week, so that to predict the load of a future day, it is necessary to perform daily segmentation on a weekly segmentation basis.
(2) Construction of a reference load Curve
First, 7 daily load subsets in step one are set as S = [ S ] 1 ,s 2 ,s 3 ,s 4 ,s 5 ,s 6 ,s 7 ]These seven subsets correspond to all daily load history data for monday through 7 days of sunday, respectively. Then, the average value of all subsets is calculated respectively to obtain
M＝[m 1 ,m 2 ,m 3 ,m 4 ,m 5 ,m 6 ,m 7 ] (1)
Wherein m is j Is s j J =1, \8230;, 7, using the average as a typical load curve for 7 days per day on monday to sunday.
Load data N days before the date to be predicted is then selected as input for constructing a reference load curve. It is defined as:
H＝[h 1 ,h 2 ,…h i …,h N ] (2)
suppose h N+1 Is the load data of the predicted day, and the formula (2) represents the historical load data of the previous N days of the predicted day, i =1, \8230;
and then, finding out the most similar typical load curve m corresponding to each h through cosine similarity.
The calculation procedure is as follows.
C in formula (3) i Is h i The corresponding most similar typical load curve index, i.e. indicate h i And
most similar. .
d i ＝(c i +N-i+1)mod7 (4)
Wherein d is i Representation according to historical data h i ，
Most likely occurring on day N + 1. Thus, for a plurality of h i In h with i And
the similarity of the data is used as a weight value, and a reference load curve of the day to be predicted is calculated:
wherein the content of the first and second substances,
is the final reference load curve, d i ＝(c i + N-i + 1) mod7 denotes a serial number,
denotes the d-th i Typical load curve corresponding to day, i.e. according to historical data h i ，
Most likely occurring on day N + 1. N represents a historical load data sharing N days.
(3) Correcting a reference load curve using a prediction model
Based on the residual error idea, the invention designs a prediction model to predict the difference value between the actual load and the reference load curve, and then corrects the reference load curve to obtain the final prediction result.
The invention designs a prediction model combining a time convolution network with a double attention mechanism. The network calculation formula of the self-attention mechanism and the external attention mechanism is as follows:
r＝E v (Norm(E k (x)) (7)
where n is the dimension of x. The training formula of the time convolution network is as follows:
h＝[h 1 ,h 2 ,…,h n ] (9)
s＝W·(h+(z+r)) (10)
wherein w is an element of z + r. f (-) is a filter and k is its dimension. d is a spreading factor.
In order to increase the stability and reliability of model prediction and reduce the sensitivity of the model at the optimal result, the invention adopts Mean Square Error (MSE) as an accuracy index and selects proper N in combination with training time so as to realize the balance between accuracy and time cost. The mean square error is calculated as follows:
(4) Obtaining a predicted result
And (3) taking the historical load data as a training set, sequentially taking data from the historical load sequence by adopting a sliding window method, inputting the data into the network for training, and determining a piece of historical data with a proper length as an input after the network is trained, namely determining N in the formula (2).
And after the trained network is obtained, inputting historical load data of N days before the predicted day into the network to obtain output, and adding the output and the reference load curve to obtain the load predicted value of the next day.
It will be understood by those skilled in the art that the foregoing is only an exemplary embodiment of the present invention, and is not intended to limit the invention to the particular forms disclosed, since various modifications, substitutions and improvements within the spirit and scope of the invention are possible and within the scope of the appended claims.
Claims (7)
1. A method for short-term prediction of residential load, comprising:
s1, performing variational modal decomposition on historical load data subjected to data cleaning to obtain a training set;
s2, performing primary segmentation on the historical load data subjected to data cleaning according to natural weeks, averaging all subsequences subjected to primary segmentation, and performing secondary segmentation on the average sequence by taking days as a unit;
s4, taking the average value sequence as a typical load curve of each day, selecting load data N days before the date to be predicted, and calculating by taking the similarity between the load data N days before and the typical load curve as a weight to obtain a reference load curve of the date to be predicted;
s5, sequentially fetching data from the training set by adopting a sliding window method and inputting the data into a prediction network for iterative training to obtain a trained prediction model; the prediction network comprises a self-attention mechanism, an external attention mechanism and a time convolution network;
the self-attention mechanism is used for distributing weights to the input data according to the internal features of the input data to enable the internal key feature weight to be increased and the weights of other features to be decreased so as to help the prediction model to extract key features;
an external attention mechanism, which is used for distributing weight to the current input data according to the characteristics of all the input data in the training set, so that the periodically related characteristic weight is increased, and the other characteristic weights are decreased;
the time convolution network is used for learning the data processed by the self-attention mechanism and the external attention mechanism, changing the value of an internal parameter through error back propagation, extracting the periodicity and the time dependency of the data and outputting a prediction result;
and S6, inputting the historical load data N days before the measured day into the trained prediction model to obtain a load prediction result, and overlapping the load prediction result with a reference load curve to obtain a load prediction value of the next day.
2. The method according to claim 1, wherein the data cleansing comprises filling in nulls in the data and removing data with a load less than a threshold for a set duration.
3. The method for short-term prediction of residential load according to claim 2, characterized in that the reference load curve is calculated by the formula:
4. A system for short term prediction of residential load, comprising:
the first data preprocessing module is used for performing variation modal decomposition on the historical load data subjected to data cleaning to serve as a training set;
the second data preprocessing module is used for carrying out primary segmentation on the historical load data subjected to data cleaning according to natural weeks, solving the average value of all subsequences of the primary segmentation, and carrying out secondary segmentation on the average value sequence by taking days as a unit;
the reference coincidence curve construction module is used for taking the average value sequence as a typical load curve of each day, selecting load data N days before the date to be predicted, and calculating the similarity between the load data N days before and the typical load curve as a weight to obtain a reference load curve of the date to be predicted;
the prediction model training module is used for sequentially fetching data from the training set by adopting a sliding window method and inputting the data into a prediction network for iterative training to obtain a trained prediction model; the prediction network comprises a self-attention mechanism, an external attention mechanism and a time convolution network;
the self-attention mechanism is used for distributing weights to input data according to internal features of the input data to enable the internal key feature weight to be increased and the weights of other features to be reduced, and helps the prediction model to extract key features;
the external attention mechanism is used for distributing weight to the current input data according to the characteristics of all input data in the training set, so that the periodically related characteristic weight is increased, and the other characteristic weights are reduced;
the time convolution network is used for learning the data processed by the self-attention mechanism and the external attention mechanism, changing the value of an internal parameter through error back propagation, extracting the periodicity and the time dependency of the data and outputting a prediction result;
and the short-term load prediction module is used for inputting the historical load data of N days before the measured day into the trained prediction model to obtain a load prediction result, and superposing the load prediction result and a reference load curve to obtain a load prediction value of the next day.
5. The residential load short-term prediction system as claimed in claim 4, wherein the data cleansing comprises filling in nulls in the data and removing data with a load less than a certain threshold for a set period of time.
6. The system for short term prediction of residential load as claimed in claim 5, wherein the reference load curve is calculated by the formula:
7. A computer-readable medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211377095.XA CN115689026A (en) | 2022-11-04 | 2022-11-04 | Method and system for short-term prediction of house load |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211377095.XA CN115689026A (en) | 2022-11-04 | 2022-11-04 | Method and system for short-term prediction of house load |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115689026A true CN115689026A (en) | 2023-02-03 |
Family
ID=85049648
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211377095.XA Pending CN115689026A (en) | 2022-11-04 | 2022-11-04 | Method and system for short-term prediction of house load |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115689026A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117808175A (en) * | 2024-03-01 | 2024-04-02 | 南京信息工程大学 | Short-term multi-energy load prediction method based on DTformer |
-
2022
- 2022-11-04 CN CN202211377095.XA patent/CN115689026A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117808175A (en) * | 2024-03-01 | 2024-04-02 | 南京信息工程大学 | Short-term multi-energy load prediction method based on DTformer |
| CN117808175B (en) * | 2024-03-01 | 2024-05-17 | 南京信息工程大学 | DTformer-based short-term multi-energy load prediction method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110610280B (en) | Short-term prediction method, model, device and system for power load | |
| Beyca et al. | Using machine learning tools for forecasting natural gas consumption in the province of Istanbul | |
| Aprillia et al. | Statistical load forecasting using optimal quantile regression random forest and risk assessment index | |
| Hyndman et al. | Density forecasting for long-term peak electricity demand | |
| Todini | A model conditional processor to assess predictive uncertainty in flood forecasting | |
| Stankovic et al. | A graph-based signal processing approach for low-rate energy disaggregation | |
| CN112488396A (en) | Wavelet transform-based electric power load prediction method of Holt-Winters and LSTM combined model | |
| Zhou et al. | Predicting energy consumption: A multiple decomposition-ensemble approach | |
| Zhang et al. | A novel power‐driven grey model with whale optimization algorithm and its application in forecasting the residential energy consumption in China | |
| CN112215442A (en) | Method, system, device and medium for predicting short-term load of power system | |
| CN112396225A (en) | Short-term electricity price prediction method based on long-term and short-term memory network | |
| Martínez‐Rodríguez et al. | Particle swarm grammatical evolution for energy demand estimation | |
| He et al. | A cooperative ensemble method for multistep wind speed probabilistic forecasting | |
| CN115689026A (en) | Method and system for short-term prediction of house load | |
| CN112365056A (en) | Electrical load joint prediction method and device, terminal and storage medium | |
| CN111985719A (en) | Power load prediction method based on improved long-term and short-term memory network | |
| CN116054156A (en) | Smart power grid short-term load prediction method, smart power grid short-term load prediction system and storage medium | |
| Zhang et al. | Accurate ultra-short-term load forecasting based on load characteristic decomposition and convolutional neural network with bidirectional long short-term memory model | |
| CN116777066B (en) | Photovoltaic power probability prediction method and device based on foundation cloud image guided depth decomposition | |
| CN112421608A (en) | Family load prediction method based on Markov from bottom to top | |
| Raman et al. | Demand baseline estimation using similarity‐based technique for tropical and wet climates | |
| CN116885705A (en) | Regional load prediction method and device | |
| CN115619447A (en) | Monthly electricity sales combined prediction method, equipment and medium | |
| CN114925940A (en) | Holiday load prediction method and system based on load decomposition | |
| CN115217152A (en) | Method and device for predicting opening and closing deformation of immersed tunnel pipe joint |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |