CN110866645A - Ultra-short-term load prediction method and system based on deep learning - Google Patents
Ultra-short-term load prediction method and system based on deep learning Download PDFInfo
- Publication number
- CN110866645A CN110866645A CN201911120087.5A CN201911120087A CN110866645A CN 110866645 A CN110866645 A CN 110866645A CN 201911120087 A CN201911120087 A CN 201911120087A CN 110866645 A CN110866645 A CN 110866645A
- Authority
- CN
- China
- Prior art keywords
- prediction
- data
- model
- analysis
- ultra
- 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
- 238000013135 deep learning Methods 0.000 title claims abstract description 39
- 238000000034 method Methods 0.000 title claims abstract description 23
- 238000012549 training Methods 0.000 claims abstract description 67
- 238000004458 analytical method Methods 0.000 claims abstract description 56
- 238000012360 testing method Methods 0.000 claims abstract description 36
- 238000007619 statistical method Methods 0.000 claims abstract description 24
- 238000010248 power generation Methods 0.000 claims abstract description 20
- 230000009286 beneficial effect Effects 0.000 claims abstract description 13
- 238000012800 visualization Methods 0.000 claims abstract description 13
- 230000000007 visual effect Effects 0.000 claims abstract description 12
- 238000005457 optimization Methods 0.000 claims abstract description 9
- 230000000694 effects Effects 0.000 claims description 9
- 238000005065 mining Methods 0.000 claims description 6
- 238000012163 sequencing technique Methods 0.000 claims description 6
- 238000013527 convolutional neural network Methods 0.000 claims description 5
- 238000007781 pre-processing Methods 0.000 claims description 5
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 4
- 230000006870 function Effects 0.000 claims description 4
- 238000010921 in-depth analysis Methods 0.000 claims description 4
- 230000007774 longterm Effects 0.000 claims description 4
- 230000002349 favourable effect Effects 0.000 claims description 2
- 238000005516 engineering process Methods 0.000 description 6
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000001965 increasing effect Effects 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000011835 investigation Methods 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000003042 antagnostic effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000002085 persistent effect Effects 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 230000000630 rising effect Effects 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 239000009891 weiqi Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E40/00—Technologies for an efficient electrical power generation, transmission or distribution
- Y02E40/70—Smart grids as climate change mitigation technology in the energy generation sector
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Strategic Management (AREA)
- Human Resources & Organizations (AREA)
- Software Systems (AREA)
- Tourism & Hospitality (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- General Business, Economics & Management (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- Marketing (AREA)
- Life Sciences & Earth Sciences (AREA)
- Primary Health Care (AREA)
- Water Supply & Treatment (AREA)
- Public Health (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides an ultra-short term load prediction method based on deep learning, which is used for accurately predicting 15-minute ultra-short term load data. In the actual power generation load prediction data of the whole province in Hunan province, the ultra-short term load prediction method provided by the invention can achieve 99.43% of accuracy on a training set and a test set. And powerful decision basis can be provided for the actual power generation planning. Meanwhile, on the basis of the prediction accuracy index, statistical analysis of a prediction result, visualization analysis of a predicted value and analysis of a prediction deviation time point are provided. The statistical analysis of the prediction results is used for providing a statistical confidence basis for the reliability of the model prediction after training. And the predicted value visual analysis is used for visually displaying the difference between the predicted value of the prediction model and the actual power load prediction data. The forecast deviation time point analysis is used for providing analysis for the time point of the value with the big forecast deviation, and is beneficial to the power generation condition optimization of later-stage power grid workers in specific time.
Description
Technical Field
The invention relates to the field of load prediction of a power system, in particular to an ultra-short-term load prediction method and system based on deep learning.
Background
The power industry is the fundamental industry of national economy. Along with the improvement of the industrial structure in China and the whole living standard of people, the demand on electric energy is increased year by year, the requirement on the electric power quality is higher and higher, and higher requirements are provided for the construction and layout of a power grid due to the simultaneity of the electric energy production and consumption. The power load prediction is the basis and foundation for planning and constructing the power grid. With the electric power industry playing an increasingly important role in national economy, correct prediction of power load is of particular importance. The power load prediction refers to that the internal connection and the development change rule among things are explored by analyzing and researching historical load data of a power system and applying qualitative and quantitative methods such as statistics, mathematics, computers, engineering techniques, empirical analysis and the like, and the future load development is estimated and speculated in advance. The accuracy of the power load prediction result directly relates to the benefits of power investment, the reliability of power supply, the normal development of power utilization requirements, and social and economic benefits. However, it is difficult to make the prediction accurate or more accurate, because there are many factors that affect the power load prediction, and because the industrial structure and the living standard of people are different in each region, the sensitivity of each specific factor to the power load prediction is different, so the power load prediction is ambiguous.
In recent years, Deep learning (Deep learning) related technologies have achieved enormous achievements in various fields. In 2012, deep learning techniques were applied to ImageNet image recognition competitions and made deep learning the focus of the public by rolling the second best performance. In 2013, the deep learning technology is applied to the face recognition technology, and good results are obtained; in 2015, the precision of the deep learning technology on the ImageNet image recognition match exceeds the precision of human eye recognition; in 2016, Google develops an Alpha Go program based on a deep learning algorithm and defeats the world champion frizzy of the Weiqi; in 2018, a convolutional neural network and an antagonistic generation network in a deep learning technology are used for improving the super-resolution ratio of pictures, enhancing data, migrating styles and other fields. The convolutional neural network in deep learning has an extremely excellent capability for image processing, and also has great achievement in the fields of text processing, voice recognition and the like. At present, the research on the structure and theory of the convolutional neural network is still a hot spot in the field of deep learning research.
An effective solution to the problems in the related art has not been proposed yet.
Disclosure of Invention
Aiming at the technical problems in the related art, the invention provides an ultra-short-term load prediction method based on deep learning, and aims to solve the problem of fuzzy prediction precision in ultra-short-term load prediction and provide statistical analysis, visualization analysis and prediction deviation analysis for a prediction result.
In order to achieve the technical purpose, the technical scheme of the invention is realized as follows:
the invention discloses an ultra-short term load prediction system based on deep learning, which comprises:
the characteristic engineering is used for carrying out characteristic mining on the original power load prediction data and extracting characteristics beneficial to final prediction so as to improve the ultra-short-term load precision;
the prediction model is used for performing fitting training on power load prediction after characteristic engineering, and the trained model comprises a learned load prediction rule;
the statistical module is used for carrying out statistical analysis on the prediction result and providing a reliability data result on statistical learning for the reliability of the model prediction after training;
the predicted value visual analysis module is used for visually displaying the curve difference between the predicted value of the prediction model and the actual power load prediction data;
and the prediction deviation time point analysis module is used for providing analysis for the time point of the value with the large prediction deviation, and is favorable for power generation data conditions in specific time to be optimized by power grid workers in the later period.
Further, a time window is selected and first derivative of past time points is selected, wherein the time window is to select data points of two past days, and for each time point, the closest 4 points to each time point are selected for first derivative.
Furthermore, the prediction model is composed of a convolutional neural network and an artificial neural network, wherein the prediction model comprises 6 layers, the first layer is a convolutional layer, and the rest are all connected layers; the features of the load expected data on the context space are extracted through the convolutional layer, and the features extracted by the convolutional layer are combined continuously through the full-connection layer, so that a good precision prediction effect is tested.
The loss function of the model is the Mean Absolute Error (MAE), which is specifically calculatedThe formula is
When the model is trained, the lower the MAE value on the test set is, the better the training effect of the model is, namely the rule of the model learning the load prediction data is obtained;
the final prediction precision of the model is calculated according to the formula
Wherein, higher final accuracy of the calculation indicates better final effect.
Still further, statistically analyzing the predicted outcomes of the test set further comprises: on the basis of the average absolute error and the prediction precision of the predicted result, expanding the standard deviation of the prediction error of the maximum value, the minimum value and the mean value of the prediction error, wherein the maximum value of the prediction error represents the prediction condition under the least ideal condition, the minimum value of the prediction error represents the optimal prediction condition, the mean value of the prediction error represents the average prediction level, and the standard deviation of the prediction error represents the fluctuation condition of the overall prediction result; thereby providing a statistical analysis of the predicted outcome.
Still further, the predictive value visualization analysis further comprises: through the visualized actual load prediction data and the model predicted load prediction data curve, the time period data prediction deviation is intuitively observed to be larger, so that secondary optimization can be conveniently carried out in the later period.
Still further, the predictive deviation time point analysis further comprises: and predicting the test set, sorting the prediction deviation from large to small, then taking the data points of the top 10% with the largest prediction deviation, and visualizing the data points with larger prediction deviation of the data at which time point of each day.
The invention further discloses an ultra-short term load prediction method based on deep learning, which comprises the following steps:
step 1: acquiring an original data set, visualizing the data set, and observing whether the data set has a certain rule in a short-term, medium-term and long-term range;
step 2: preprocessing data, and sequencing the obtained original data from small to large according to the time sequence;
and step 3: performing feature mining on the original power load prediction data, and extracting features beneficial to final prediction so as to improve the ultra-short-term load precision;
and 4, step 4: and constructing a prediction model and training the model, wherein the prediction model is a mixed model of a convolutional layer and a full-connection layer with 6 layers.
And 5: and when the model is trained on the training set, verifying the accuracy of the model on the testing set.
Further, the step 2 specifically includes: sequencing all the data according to a time sequence to form a one-dimensional data format; then selecting the range of a period of time in the past as the input of the model, and taking the time point to be predicted in the future as the output, namely the training label of the model; finally, the window is continuously slid to get our final total data set, where the data window size is set to a hyper-parameter setting, and when the historical data window is selected to be 2 days, there are 192 data points.
Further, the step 3 specifically includes: for each data of the time points of two past days, the first derivative of the nearest 4 adjacent points is taken, and the shape of the processed single input data is as follows: 5 x 193; all training data are not scattered randomly, but are divided according to the time sequence, wherein the first 80% of the data set is used as a training set, and the last 20% of the data set is used as a testing set.
Further, the step 4 specifically includes: after the model is built and all data sets are processed, the model is trained; dividing a training set and a test set according to a time sequence; during training, all of the training sets are broken up, with each training batch having a size of 128.
Further, the step 5 specifically includes: when the model is trained on the training set, the accuracy of the model on the testing set is verified, and meanwhile, in-depth analysis is carried out on the prediction deviation, wherein the in-depth analysis comprises prediction result statistical analysis, prediction value visualization analysis and prediction deviation time point analysis.
The invention has the beneficial effects that: the ultra-short term load prediction method based on deep learning according to the above embodiments of the present invention can provide accurate prediction for ultra-short term load prediction of 15 minutes. On the actual power generation load prediction data of the whole province in Hunan province, 99.43% accuracy can be achieved on the training set and the testing set by training according to the steps. And a powerful decision basis can be provided for the actual power generation planning. Meanwhile, on the basis of the prediction accuracy index, statistical analysis of a prediction result, visualization analysis of a predicted value and analysis of a prediction deviation time point are provided. The prediction result statistical analysis is used for providing a reliability data result on statistical learning for the reliability of the model prediction after training. And the predicted value visual analysis is used for visually displaying the curve difference between the predicted value of the prediction model and the actual power load prediction data. The forecast deviation time point analysis is used for providing analysis for the time point of the value with the big forecast deviation, and is beneficial to the power generation data condition optimization of later-stage power grid workers in specific time.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.
FIG. 1 is a schematic structural diagram of an ultra-short term load prediction model based on deep learning according to the present invention;
fig. 2 is a flowchart of the ultra-short term load prediction method based on deep learning according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments that can be derived by one of ordinary skill in the art from the embodiments given herein are intended to be within the scope of the present invention.
Example one
The invention provides an ultra-short-term load prediction method based on deep learning, and aims to solve the problem of fuzzy prediction precision in ultra-short-term load prediction and provide statistical analysis, visualization analysis and prediction deviation analysis for prediction results.
And carrying out visual analysis on the original data. The obtained original data set is subjected to visual analysis, and the purpose is to visually display the volatility and regularity of the original data set. Specifically, we can visualize 3 days, 3 months and the entire data set, respectively. We found that the data set exhibited strong regularity at 96 data points a day, but that the regularity was decreasing with increasing time. And the power generation of the annual average data point shows a rising trend with the increase of time. Among them, the average time point of the entire province in Hunan in 2017 was 14818, the number of the province in 2018 increased to 16859, and the average time point of the province in the 4 months before 2019 was 17829. The reason why the power generation amount at this average time point per year tends to increase is considered to be that the average power generation amount per year tends to increase gradually as the economy progresses. And the monthly power generation trends of 2017 and 2018 have certain correlation. But at the same time, this also creates a difficult prediction: the power generation amount at 2019 is higher than the power generation amount at 2017 and 2018, which indicates that the form of data of 2019 never appears in the past. This can create great difficulty for ultra-short term load prediction.
And (4) preprocessing data. After intuitive analysis of the raw data, it is the data format that we need to process the data set into post model training. In specific data processing, all data are sequenced according to time sequence to form a one-dimensional data format. Then, we select the range of a past time point as the input of the model, and the time point to be predicted in the future as the output, i.e. the training label of the model. Finally, we continuously slide the window to get our final full dataset. Wherein the data window size is set to a hyper-parameter setting. Experimental comparisons were required to obtain optimal results. We found that the prediction was best when the historical window of data was selected for a total of 192 data points over 2 days.
And (5) characteristic engineering. The characteristic engineering is used for carrying out characteristic mining on the original power load prediction data and extracting characteristics beneficial to final prediction so as to improve the ultra-short-term load precision. In the application of the deep learning technology, the feature engineering plays an important role in ultra-short-term load prediction and is directly related to the final result of prediction accuracy. We take the first derivative of the nearest 4 neighbors to each data at the time points of the last two days. The shape of the processed single input data is as follows: 5 x 193; in this example, all the training data is not scattered randomly, but divided in time order. We treated the first 80% of the data set as the training set and the last 20% as the test set. The final number of training sets is 63664 and the number of test sets is 15916.
And predicting the model structure. The specific configuration of the model structure is shown in fig. 1, and the model includes a total of 6 input layers. Wherein the input data format is 5 × 193; the first layer of the hidden layers is the convolutional layer, the convolution operation type is Conv 1D. The number of convolution kernels is 32, and the step size is 1; after the memorable one-dimensional convolution, the output data needs to be straightened for the subsequent full join operation. The types of the hidden layers of the 2 nd, the 3 rd and the 4 th layers are all connected, and the number of the neurons is 256. The last layer is an output layer, and the number of neurons is only 1 because the load prediction data is to be predicted. All activation functions are Relu throughout the model. After the model was built, all training parameters for the model were 1673025. The optimizer for model training is Adam, and the learning rate is 0.001 by default.
And (5) training a model. After the model is built and all data sets are processed, training of the model is started. For the division of the training set and the test set, we divide in time order. However, during the training process, we break up all training sets. Where the size of each trained batch is 128. The setting of the size of batch is also a hyper-parameter, which affects the final prediction accuracy. Some more exotic phenomena sometimes occur during training. That is, the Loss of the model is not substantially reduced when the model is initially trained. We hypothesize that this may be the result of initializing model parameters, but this poor phenomenon is not always persistent. If the above situation occurs, the training is stopped and then continued. After training, the Loss of the model on the training set is 83.3592, and the Loss on the testing set is 96.6481; the final prediction accuracy on the training set was calculated to be 0.9943 and the training accuracy on the test set was calculated to be 0.9943. This indicates that the model is able to learn well about the laws contained in the data set.
And (5) carrying out statistical analysis on the prediction result. And the statistical analysis of the prediction results is mainly to perform statistical analysis on the prediction results of the test set. Relying solely on the mean absolute error may miss some of the wanted information at the end of the training. Therefore, the standard deviation of the prediction errors of the maximum value, the minimum value and the mean value of the prediction errors is expanded on the basis of the average absolute error and the prediction precision of the predicted result. The maximum value of the prediction error represents the prediction condition under the condition of least ideal, the minimum value of the prediction error represents the optimal prediction condition, the mean value of the prediction error represents the average prediction level, and the standard deviation of the prediction error represents the fluctuation condition of the whole prediction result; in this example, the model predicted a minimum value of 0.0166, a maximum value of 723.3066 on the test set in the Hunan province, an average prediction deviation of 96.8388 on the test set, and a standard deviation of the prediction deviation on the test set of 81.1503. This indicates that the final prediction results have a high confidence. The predicted deviation averages fluctuate within 81.1503 above and below the predicted deviation averages.
And (5) visually analyzing the predicted value. The visual analysis of the predicted value can provide the most visual predicted result for people. By visualizing the actual load prediction data and the model predicted load prediction data curve, the larger data prediction deviation in the time period can be intuitively observed, so that the secondary optimization can be conveniently carried out in the later period.
And (5) analyzing the predicted deviation time points. Predictive deviation time point analysis provides that the data point for which the predictive deviation is greater occurs at that time point of the day. Specifically, we predict the test set and sort the prediction deviations from large to small. We take the first 10% of the data points with the largest prediction deviation and visualize the data points where the larger prediction deviation of these data occurs at that time point of the day. The analysis is beneficial to the power generation data situation of later-period power grid workers in specific time to carry out actual investigation and optimization, and the situation is not required to be observed in 24 hours all day.
The ultra-short term load prediction method based on deep learning according to the above embodiments of the present invention can provide accurate prediction for ultra-short term load prediction of 15 minutes. On the actual power generation load prediction data of the whole province in Hunan province, 99.43% accuracy can be achieved on the training set and the testing set by training according to the steps. And a powerful decision basis can be provided for the actual power generation planning. Meanwhile, on the basis of the prediction accuracy index, statistical analysis of a prediction result, visualization analysis of a predicted value and analysis of a prediction deviation time point are provided. The prediction result statistical analysis is used for providing a reliability data result on statistical learning for the reliability of the model prediction after training. And the predicted value visual analysis is used for visually displaying the curve difference between the predicted value of the prediction model and the actual power load prediction data. The forecast deviation time point analysis is used for providing analysis for the time point of the value with the big forecast deviation, and is beneficial to the power generation data condition optimization of later-stage power grid workers in specific time.
The method for predicting the ultra-short-term load based on deep learning comprises the following specific implementation steps:
step 1: the raw data set is acquired and visualized. Observe whether the data set has certain regularity in the short, medium and long term ranges.
Step 2: and (4) preprocessing data. And sequencing the acquired original data from small to large according to the time sequence. Then, the time points of the last two days are taken as a window, and the time point data corresponding to the ultra-short term prediction time required by the time is taken as a label. All data sets are obtained through a continuous sliding window.
And step 3: and (5) characteristic engineering. The characteristic engineering plays an important role in ultra-short-term load prediction and is directly related to the final prediction precision result. We take the first derivative of the nearest 4 neighbors to each data at the time points of the last two days. In this example, all the training data is scattered randomly, but divided in time order. We treated the first 80% of the data set as the training set and the last 20% as the test set.
And 4, step 4: and (5) training a model. In this example, we built a hybrid model of 6 convolutional layers and fully-connected layers. The input layer and the output layer are removed, and a total of 4 hidden layers are provided. Wherein the first layer of the hidden layer is a convolution layer, and the other 3 layers are all connection layers. The optimizer of the model is Adam, and the learning rate is 0.001 by default.
And 5: model prediction and prediction bias analysis. When the model is trained on the training set, it is not enough to obtain higher precision on the training set, and it is also necessary to verify the precision on the test set. Meanwhile, in order to perform deep analysis on the prediction deviation, statistical analysis of prediction results, visualization analysis of predicted values and analysis of prediction deviation time points are also required.
The programming language, deep learning framework and other related dependency packages involved in the present example are: the programming language is Python 3.6; the deep learning framework is Keras developed by Google, and the back-end implementation is Tensorflow; visualization related dependency packages are mainly Matplotlib; all the third-party packages involved in the embodiment of the invention are open sources, and no property dispute exists.
Example two
As shown in fig. 2, the ultra-short term load prediction method based on deep learning according to the embodiment of the present invention aims to solve the problem of fuzzy prediction accuracy when ultra-short term load prediction is performed, and provide statistical analysis, visualization analysis and prediction bias analysis for the prediction result.
In order to achieve the above object, an embodiment of the present invention provides a method for predicting an ultra-short term load based on deep learning, including:
and the characteristic engineering is used for carrying out characteristic mining on the original power load prediction data and extracting characteristics beneficial to final prediction so as to improve the ultra-short-term load precision.
And the prediction model is used for performing fitting training on the power load prediction after the characteristic engineering, and the trained model comprises the learned load prediction rule.
And the prediction result statistical analysis is used for providing a reliability data result in statistical learning for the reliability of the model prediction after training.
And the visual analysis of the predicted value is used for visually displaying the curve difference between the predicted value of the prediction model and the actual power load prediction data.
And the prediction deviation time point analysis is used for providing analysis for the time point of the value with the large prediction deviation, and is beneficial to the later-stage power grid staff to optimize the power generation data condition in the specific time.
Wherein the feature engineering comprises selection of a time window for ultra-short term load anticipation and a first derivative of past time points. Through a hyper-parametric comparative test, the prediction precision of the data points of the last two days is the highest. Meanwhile, for each time point, the prediction accuracy of the first derivative of the 4 points closest to the time point is the highest through hyper-parametric experiments.
The prediction model is composed of a convolution neural network and an artificial neural network. Specifically, the prediction model has 6 layers, wherein the first layer is a convolution layer, and the rest are all connection layers. The features of the load expected data on the context space are extracted through the convolutional layer, and the features extracted by the convolutional layer are combined continuously through the full-connection layer, so that a good precision prediction effect is tested.
Loss function of modelThe number is the Mean Absolute Error (MAE), which can reflect the actual situation of the Error of the predicted value well. The specific calculation formula is
When the model is trained, the lower the MAE value on the test set is, the better the training effect of the model is considered, namely the model well learns the rule of load prediction data;
the final prediction precision of the model is calculated according to the formula
Higher final accuracy of the calculation indicates better final effect.
The statistical analysis of the prediction results is mainly to perform statistical analysis on the prediction results of the test set. Relying solely on the mean absolute error may miss some of the wanted information at the end of the training. Therefore, the standard deviation of the prediction errors of the maximum value, the minimum value and the mean value of the prediction errors is expanded on the basis of the average absolute error and the prediction precision of the predicted result. The maximum value of the prediction error represents the prediction condition under the condition of least ideal, the minimum value of the prediction error represents the optimal prediction condition, the mean value of the prediction error represents the average prediction level, and the standard deviation of the prediction error represents the fluctuation condition of the whole prediction result; by providing statistical analysis of the prediction results, one can gain a more complete understanding of the integrity of the prediction data.
The visual analysis of the predicted value can provide the most visual predicted result for people. By visualizing the actual load prediction data and the model predicted load prediction data curve, the larger data prediction deviation in the time period can be intuitively observed, so that the secondary optimization can be conveniently carried out in the later period.
Wherein the prediction deviation time point analysis provides the data point at which the greater value of the prediction deviation occurs at each time point. Specifically, we predict the test set and sort the prediction deviations from large to small. We take the first 10% of the data points with the largest prediction deviation and visualize the data points where the larger prediction deviation of these data occurs at that time point of the day. The analysis is beneficial to the power generation data situation of later-period power grid workers in specific time to carry out actual investigation and optimization, and the situation is not required to be observed in 24 hours all day.
The embodiment of the invention also provides specific algorithm steps of the ultra-short term load prediction method based on deep learning, which comprise the following steps:
step 1: the raw data set is acquired and visualized. Observe whether the data set has certain regularity in the short, medium and long term ranges.
Step 2: and (4) preprocessing data. And sequencing the acquired original data from small to large according to the time sequence. Then, the time points of the last two days are taken as a window, and the time point data corresponding to the ultra-short term prediction time required by the time is taken as a label. All data sets are obtained through a continuous sliding window.
And step 3: and (5) characteristic engineering. The characteristic engineering plays an important role in ultra-short-term load prediction and is directly related to the final prediction precision result. We take the first derivative of the nearest 4 neighbors to each data at the time points of the last two days. In this example, all the training data is scattered randomly, but divided in time order. We treated the first 80% of the data set as the training set and the last 20% as the test set.
And 4, step 4: and (5) training a model. In this example, we built a hybrid model of 6 convolutional layers and fully-connected layers. The input layer and the output layer are removed, and a total of 4 hidden layers are provided. Wherein the first layer of the hidden layer is a convolution layer, and the other 3 layers are all connection layers. The optimizer of the model is Adam, and the learning rate is 0.001 by default.
And 5: model prediction and prediction bias analysis. When the model is trained on the training set, it is not enough to obtain higher precision on the training set, and it is also necessary to verify the precision on the test set. Meanwhile, in order to perform deep analysis on the prediction deviation, statistical analysis of prediction results, visualization analysis of predicted values and analysis of prediction deviation time points are also required.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (10)
1. An ultra-short term load prediction system based on deep learning, comprising:
the characteristic engineering is used for carrying out characteristic mining on the original power load prediction data and extracting characteristics beneficial to final prediction so as to improve the ultra-short-term load precision;
the prediction model is used for performing fitting training on power load prediction after characteristic engineering, and the trained model comprises a learned load prediction rule;
the statistical module is used for carrying out statistical analysis on the prediction result and providing a reliability data result on statistical learning for the reliability of the model prediction after training;
the predicted value visual analysis module is used for visually displaying the curve difference between the predicted value of the prediction model and the actual power load prediction data;
and the prediction deviation time point analysis module is used for providing analysis for the time point of the value with the large prediction deviation, and is favorable for power generation data conditions in specific time to be optimized by power grid workers in the later period.
2. The ultra-short term load prediction system based on deep learning as claimed in claim 1, wherein the time window is selected and first derivative of past time points is selected, wherein the time window is selected for two days in the past, and for each time point, the closest 4 points to each time point are selected for first derivative.
3. The ultra-short term load prediction system based on deep learning of claim 2, wherein the prediction model is composed of a convolutional neural network and an artificial neural network, wherein the prediction model comprises 6 layers, wherein the first layer is a convolutional layer, and the rest are all connected layers; extracting the characteristics of the load expected data on the context space through the convolutional layer, and continuously combining the characteristics extracted by the convolutional layer through the fully-connected layer so as to test a better precision prediction effect;
the loss function of the model is the Mean Absolute Error (MAE) which is specifically calculated by the formula
When the model is trained, the lower the MAE value on the test set is, the better the training effect of the model is, namely the rule of the model learning the load prediction data is obtained;
the final prediction precision of the model is calculated according to the formula
Wherein, higher final accuracy of the calculation indicates better final effect.
4. The deep learning-based ultra-short term load prediction system of claim 3, wherein statistically analyzing the prediction results of the test set further comprises: on the basis of the average absolute error and the prediction precision of the predicted result, expanding the standard deviation of the prediction error of the maximum value, the minimum value and the mean value of the prediction error, wherein the maximum value of the prediction error represents the prediction condition under the least ideal condition, the minimum value of the prediction error represents the optimal prediction condition, the mean value of the prediction error represents the average prediction level, and the standard deviation of the prediction error represents the fluctuation condition of the overall prediction result; thereby providing a statistical analysis of the predicted outcome.
5. The deep learning-based ultra-short term load prediction system of claim 4, wherein the predictor visualization analysis further comprises: through the visualized actual load prediction data and the model predicted load prediction data curve, the time period data prediction deviation is intuitively observed to be larger, so that secondary optimization can be conveniently carried out in the later period.
6. The deep learning-based ultra-short term load prediction system of claim 5, wherein the prediction bias time point analysis further comprises: and predicting the test set, sorting the prediction deviation from large to small, then taking the data points of the top 10% with the largest prediction deviation, and visualizing the data points with larger prediction deviation of the data at which time point of each day.
7. An ultra-short term load prediction method based on deep learning is characterized by comprising the following steps:
step 1: acquiring an original data set, visualizing the data set, and observing whether the data set has a certain rule in a short-term, medium-term and long-term range;
step 2: preprocessing data, and sequencing the obtained original data from small to large according to the time sequence;
and step 3: performing feature mining on the original power load prediction data, and extracting features beneficial to final prediction so as to improve the ultra-short-term load precision;
and 4, step 4: constructing a prediction model and training the model, wherein the prediction model is a mixed model of a convolutional layer and a full-connection layer with 6 layers;
and 5: and when the model is trained on the training set, verifying the accuracy of the model on the testing set.
8. The ultra-short term load prediction method based on deep learning according to claim 7, wherein the step 2 specifically comprises: sequencing all the data according to a time sequence to form a one-dimensional data format; then selecting the range of a period of time in the past as the input of the model, and taking the time point to be predicted in the future as the output, namely the training label of the model; finally, the window is continuously slid to get our final total data set, where the data window size is set to a hyper-parameter setting, and when the historical data window is selected to be 2 days, there are 192 data points.
9. The ultra-short term load prediction method based on deep learning according to claim 7, wherein the step 3 specifically comprises: for each data of the time points of two past days, the first derivative of the nearest 4 adjacent points is taken, and the shape of the processed single input data is as follows: 5 x 193; all training data are not scattered randomly, but are divided according to the time sequence, wherein the first 80% of the data set is used as a training set, and the last 20% of the data set is used as a testing set.
10. The ultra-short term load prediction method based on deep learning according to claim 7, wherein the step 4 specifically comprises: after the model is built and all data sets are processed, the model is trained; dividing a training set and a test set according to a time sequence; during training, breaking up all the training sets, wherein the size of each training batch is 128;
and/or, the step 5 specifically comprises: when the model is trained on the training set, the accuracy of the model on the testing set is verified, and meanwhile, in-depth analysis is carried out on the prediction deviation, wherein the in-depth analysis comprises prediction result statistical analysis, prediction value visualization analysis and prediction deviation time point analysis.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911120087.5A CN110866645A (en) | 2019-11-15 | 2019-11-15 | Ultra-short-term load prediction method and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911120087.5A CN110866645A (en) | 2019-11-15 | 2019-11-15 | Ultra-short-term load prediction method and system based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110866645A true CN110866645A (en) | 2020-03-06 |
Family
ID=69653802
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911120087.5A Pending CN110866645A (en) | 2019-11-15 | 2019-11-15 | Ultra-short-term load prediction method and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110866645A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112732444A (en) * | 2021-01-12 | 2021-04-30 | 北京工业大学 | Distributed machine learning-oriented data partitioning method |
| CN116126827A (en) * | 2023-01-04 | 2023-05-16 | 三峡高科信息技术有限责任公司 | Method for simultaneously realizing modeling and visual analysis of power production data |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102982387A (en) * | 2012-10-18 | 2013-03-20 | 安徽工程大学 | Method for predicting short-term power load |
| US20130110756A1 (en) * | 2011-10-31 | 2013-05-02 | Siemens Corporation | Short-term Load Forecast Using Support Vector Regression and Feature Learning |
| CN109492836A (en) * | 2018-12-29 | 2019-03-19 | 上海电力学院 | Load forecast and Research on electricity price prediction system based on shot and long term memory network |
| CN109543901A (en) * | 2018-11-20 | 2019-03-29 | 国网辽宁省电力有限公司经济技术研究院 | Short-Term Load Forecasting Method based on information fusion convolutional neural networks model |
-
2019
- 2019-11-15 CN CN201911120087.5A patent/CN110866645A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130110756A1 (en) * | 2011-10-31 | 2013-05-02 | Siemens Corporation | Short-term Load Forecast Using Support Vector Regression and Feature Learning |
| CN102982387A (en) * | 2012-10-18 | 2013-03-20 | 安徽工程大学 | Method for predicting short-term power load |
| CN109543901A (en) * | 2018-11-20 | 2019-03-29 | 国网辽宁省电力有限公司经济技术研究院 | Short-Term Load Forecasting Method based on information fusion convolutional neural networks model |
| CN109492836A (en) * | 2018-12-29 | 2019-03-19 | 上海电力学院 | Load forecast and Research on electricity price prediction system based on shot and long term memory network |
Non-Patent Citations (2)
| Title |
|---|
| 谢德林: "《城乡电网建设改造新技术与新设备应用手册》", 31 December 1999, 煤炭工业出版社 * |
| 陆继翔等: "基于CNN-LSTM混合神经网络模型的短期负荷预测方法", 《电力系统自动化》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112732444A (en) * | 2021-01-12 | 2021-04-30 | 北京工业大学 | Distributed machine learning-oriented data partitioning method |
| CN116126827A (en) * | 2023-01-04 | 2023-05-16 | 三峡高科信息技术有限责任公司 | Method for simultaneously realizing modeling and visual analysis of power production data |
| CN116126827B (en) * | 2023-01-04 | 2023-08-04 | 三峡高科信息技术有限责任公司 | Method for simultaneously realizing modeling and visual analysis of power production data |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107578124B (en) | Short-term power load prediction method based on multilayer improved GRU neural network | |
| CN110751318A (en) | IPSO-LSTM-based ultra-short-term power load prediction method | |
| CN109615124B (en) | SCADA master station load prediction method based on deep learning | |
| CN110866645A (en) | Ultra-short-term load prediction method and system based on deep learning | |
| CN107403239B (en) | Parameter analysis method for control equipment in power system | |
| CN112257928A (en) | Short-term power load probability prediction method based on CNN and quantile regression | |
| CN117114160A (en) | Short-term photovoltaic power prediction method | |
| CN112529053A (en) | Short-term prediction method and system for time sequence data in server | |
| Yang et al. | Investigating black-box model for wind power forecasting using local interpretable model-agnostic explanations algorithm: Why should a model be trusted? | |
| AU2021106200A4 (en) | Wind power probability prediction method based on quantile regression | |
| CN115238854A (en) | Short-term load prediction method based on TCN-LSTM-AM | |
| CN109190800B (en) | Sea surface temperature prediction method based on spark frame | |
| Ruiz et al. | Predictive Energy Management in Internet of Things: Optimization of Smart Buildings for Energy Efficiency | |
| CN111090679B (en) | Time sequence data representation learning method based on time sequence influence and graph embedding | |
| Xing et al. | Predicting NVIDIA's Next-Day Stock Price: A Comparative Analysis of LSTM, MLP, ARIMA, and ARIMA-GARCH Models | |
| CN116681173A (en) | Data-intensive power load prediction parallel optimization method based on RNN model | |
| Liu et al. | Exploring Cluster Stocks based on deep learning for Stock Prediction | |
| CN111080037A (en) | Short-term power load prediction method and device based on deep neural network | |
| CN116307250A (en) | Short-term load prediction method and system based on typical daily feature selection | |
| Zhao et al. | Forecasting short-term oil price with a generalised pattern matching model based on empirical genetic algorithm | |
| Ye et al. | TADNet: Temporal Attention Decomposition Networks for Probabilistic Energy Forecasting | |
| Zheng et al. | Stock trend prediction based on ARIMA-LightGBM hybrid model | |
| Jin et al. | Time series forecasting with llms: Understanding and enhancing model capabilities | |
| CN110135624A (en) | A kind of data predication method of the combination LSTM model based on 2-D data stream | |
| Ma et al. | A Novel Hybrid Feature Importance and Feature Interaction Detection Framework for Predictive Optimization in Industry 4.0 Applications |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200306 |
|
| RJ01 | Rejection of invention patent application after publication |