CN116090604A - Training method, prediction method and device for photovoltaic power model in future and short term - Google Patents
Training method, prediction method and device for photovoltaic power model in future and short term Download PDFInfo
- Publication number
- CN116090604A CN116090604A CN202211550852.9A CN202211550852A CN116090604A CN 116090604 A CN116090604 A CN 116090604A CN 202211550852 A CN202211550852 A CN 202211550852A CN 116090604 A CN116090604 A CN 116090604A
- Authority
- CN
- China
- Prior art keywords
- data
- tcn
- photovoltaic power
- short
- channel
- 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
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000012549 training Methods 0.000 title claims abstract description 25
- 238000007781 pre-processing Methods 0.000 claims abstract description 7
- 238000010276 construction Methods 0.000 claims abstract description 6
- 230000007246 mechanism Effects 0.000 claims description 34
- 238000013528 artificial neural network Methods 0.000 claims description 14
- 238000013527 convolutional neural network Methods 0.000 claims description 12
- 238000013507 mapping Methods 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 10
- 238000004590 computer program Methods 0.000 claims description 6
- 230000002159 abnormal effect Effects 0.000 claims description 4
- 230000001364 causal effect Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 239000010410 layer Substances 0.000 description 22
- 238000010586 diagram Methods 0.000 description 7
- 230000006870 function Effects 0.000 description 6
- 238000012216 screening Methods 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000004364 calculation method Methods 0.000 description 4
- 238000010248 power generation Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 2
- 230000004913 activation Effects 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 239000002355 dual-layer Substances 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000032683 aging Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013479 data entry Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000001303 quality assessment method Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012800 visualization Methods 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
- 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)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- Economics (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Marketing (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- General Business, Economics & Management (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Evolutionary Computation (AREA)
- Tourism & Hospitality (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Primary Health Care (AREA)
- Development Economics (AREA)
- Water Supply & Treatment (AREA)
- Public Health (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Photovoltaic Devices (AREA)
Abstract
The invention relates to the technical field of solar short-term photovoltaic power prediction, in particular to a training method, a prediction method and a device of a solar short-term photovoltaic power model; after the multi-characteristic photovoltaic data are subjected to data preprocessing, the processed data set is constructed in a future prediction data set form based on time scale division; inputting the data set after division construction into a preset MHATCN-TCN model for training; the prediction accuracy of predicting the photovoltaic power of the previous short period of the future day is improved.
Description
Technical Field
The invention relates to the technical field of solar short-term photovoltaic power prediction, in particular to a training method, a prediction method and a device of a solar short-term photovoltaic power model.
Background
The photovoltaic output power has strong randomness and volatility, the uncertainty of the photovoltaic output power brings a series of scheduling operation problems, and the accurate prediction of the photovoltaic power generation output power is an effective means for reducing the influence of the uncertainty; accurate prediction of photovoltaic power plays a vital role in power system scheduling.
At present, photovoltaic power generation power prediction can be classified into ultra-short-term (intra-day) prediction, short-term (day-ahead) prediction, and medium-long-term prediction according to the predicted time scale.
The ultra-short-term power prediction is to build prediction modeling through data sources such as real-time environment monitoring data, power station inverter operation data, historical data and the like, further predict the output power of 0-4 hours in the future, and adopt a mathematical statistics method, a physical statistics and comprehensive method, and is mainly used for photovoltaic power generation power control, electric energy quality assessment and the like; such minute-level predictions typically do not employ numerical weather forecast data; the general forecast aging of the short-term (day-ahead) forecast is 0-72 hours in the future, and the numerical weather forecast is the main; the medium-long term prediction is a prediction of a long time scale, and is mainly used for maintenance arrangement of a system, prediction of power generation capacity and the like.
Aiming at a model for predicting photovoltaic power in a short period before the day, the existing model comprises CNN, TCN, CNN-LSTM, MHATCN, CNN, TCN, CNN-LSTM and MHATCN, wherein MAE of the four models is sequentially reduced, the MHATCN model shows optimal prediction precision in the four models, and the TCN model processes time sequence data; but the MHATCN model or TCN model still has low prediction accuracy when predicting the photovoltaic power of 1 day in the future for short-term photovoltaic power prediction.
Disclosure of Invention
In order to solve the defects in the prior art, the invention provides a training method, a prediction method and a device for a solar short-term photovoltaic power model, and aims to improve the prediction accuracy of predicting solar short-term photovoltaic power in the future.
A first aspect of the present invention is directed to a method for training a photovoltaic power model for a short period of time, the method comprising:
after the multi-characteristic photovoltaic data are subjected to data preprocessing, the processed data set is constructed in a future prediction data set form based on time scale division;
inputting the data set after the division construction into a preset MHATCN-TCN model for training.
As an alternative embodiment, the data start-stop time is specifically 3.5 days (336 points) and the tag start-stop time is 1 day (96 points) based on the time scale division.
As an alternative embodiment, the data preprocessing is specifically to perform maximum and minimum normalization processing on the data after processing the null value and the outlier.
Further, the construction method of the form of the future prediction data set is as follows: firstly, converting the processed data set into three-dimensional data (samples, timeps, dimensions) by using a sliding window operation to adjust the data set to the input shape required by the proposed model; the resulting data was then screened such that two adjacent samples of data were 24 hours apart.
As an alternative embodiment, the MHATCN-TCN model comprises an input layer, an MHATCN-TCN module and an output layer;
the input layer comprises multi-feature data, and all the multi-feature data are output to a first channel of the HATCN-TCN module; a separate photovoltaic output power (P) data output channel second channel;
the MHATCN-TCN module comprises a first channel and a second channel, wherein the first channel comprises a multi-head attention Mechanism (MHA) and a time convolutional neural network (TCN); the number of heads of a multi-head attention Mechanism (MHA) is the number of input features, and Q, K, V in different heads can be subjected to different spatial mapping in the multi-head attention mechanism; subsequently, taking the output of the multi-headed attentiveness Mechanism (MHA) as the input of a time convolutional neural network (TCN); the second channel adopts a double-layer TCN network; the double-layer TCN network consists of causal convolution, expansion convolution and residual error connection;
and finally, performing column splicing on the double-channel result, inputting the double-channel result into the full-connection neural network, and establishing a mapping relation between the multi-scale input and the photovoltaic power.
In the MHATCN-TCN module, the first channel convolution Kernel size is set to be Kernel_size=29; setting kernel_size=15 for the convolution Kernel size of the second channel; nb_stacks=2.
As an alternative embodiment, the multi-feature data includes photovoltaic output power (P), total irradiance (Ti), normal direct irradiance (Ni), horizontal scattered irradiance (Hi), air pressure (Ap).
A second aspect of the present invention is directed to a method for predicting photovoltaic power for a short period of time, the method comprising:
acquiring photovoltaic data of multiple characteristics to be tested;
and inputting the photovoltaic data with multiple characteristics to be tested into the MHATCN-TCN model obtained by the training method of the short-term photovoltaic power model before day in the embodiment mode to obtain the photovoltaic predicted power.
A third aspect of the invention is directed to an electronic device comprising a memory and a processor; the memory is used for storing a computer program, and the processor is used for realizing the training method of the short-term photovoltaic power model before the day or realizing the prediction method of the short-term photovoltaic power before the day when the computer program is executed.
A fourth aspect of the present invention is directed to a computer readable storage medium, on which a computer program is stored, which when executed by a processor, implements the foregoing method for training a short-term photovoltaic power model, or implements the foregoing method for predicting short-term photovoltaic power.
The beneficial effects achieved by the invention are as follows: the invention inputs the processed multi-feature data into the first channel, the first channel adopts the combination of a multi-head attention mechanism and a time convolution neural network, the number of heads of the multi-head attention is the number of input features, different space mapping is carried out on Q, K, V in different heads in the multi-head attention mechanism, and based on the design, each head possibly pays attention to different parts of input, and can represent a function more complex than a simple weighted average; q, K, V of the multi-head attention mechanism uses the same input, and its output shape is the same as the input shape by default, which facilitates the subsequent input TCN; the output of the multi-headed attention mechanism is then taken as the input to the TCN. And simultaneously, inputting the processed photovoltaic power into a second channel, wherein the second channel adopts a double-layer TCN network. TCN is composed of causal convolution, expansion convolution and residual error connection; the TCN has the advantages of parallelism and time sequence causality, and the receptive field can be flexibly adjusted, so that the TCN is very suitable for processing time sequence data, and the time sequence relativity of the photovoltaic output power can be fully mined; and finally, performing column splicing on the double-channel result and inputting the double-channel result into the fully-connected neural network. The filters of the dual-channel TCN are f1 and f2, outputs are (b, f 1) and (b, f 2), after column splicing, (b, f) and f1+f2, original data characteristics can be extracted and enriched by column splicing, and the learning capacity of a follow-up fully-connected neural network is enhanced; then inputting the model into a fully-connected neural network, and adding a Leakyrlu activation function into the fully-connected layer, wherein dropout prevents the model from being fitted excessively; the full connection can effectively retain the extracted information, and establish a mapping relation between the multi-scale input and the photovoltaic power to obtain a photovoltaic power prediction sequence.
Drawings
FIG. 1 is a schematic diagram of a data processing flow of the present invention.
Fig. 2 is a schematic diagram of the MHATCN-TCN model of the present invention.
FIG. 3 is a schematic diagram of a first channel multi-feature data entry of the present invention.
Fig. 4 is a schematic diagram of the photovoltaic output power (P) data input for the first channel of the present invention.
FIG. 5 is a graph of MHATCN-TCN model and MHATCN model predictions
Fig. 6 is a plot of MHATCN-TCN model versus MHATCN prediction scatter of the present invention.
Detailed Description
In order to facilitate understanding of the invention by those skilled in the art, a specific embodiment of the invention is described below with reference to the accompanying drawings.
The first aspect of the present application is embodied in a data set processing:
the present study uses photovoltaic field station data from a certain power station 2019, 1 month 1 day to 2020, 11 months 21 days of China, which respectively include (photovoltaic output power (P), total irradiance (Ti), normal direct irradiance (Ni), horizontal irradiance (Hi), air temperature (At), air pressure (Ap), relative humidity (Rh)), which includes a sampling interval of 15 minutes, a installed capacity of 30MW, and a total of 65534 sample data (during which there are several days of missing values), wherein 95% is a training set (644 days), and the last 5% is a test set (29 days).
Preprocessing the collected data set:
firstly, eliminating abnormal values and blank values in a data set, wherein the specific operation mode is to eliminate the abnormal values of sample data by using a box graph; then calculating an upper quartile and a lower quartile by a statistical method, and eliminating abnormal values within the range of the upper quartile and the lower quartile; finally, setting zero for the blank value through a filena function of a numpy library in python;
then, the data set is subjected to maximum and minimum normalization for easy training due to the different dimensions of different features; carrying out maximum and minimum normalization by a preprocessing method of a sklearn library in python to ensure that data in a data set is positioned in a [0,1] interval;
finally, constructing a data set form of the future prediction data set; the goal is to predict the station photovoltaic power for the next 1 day (96 points) using the 3.5 day (336 points, we want to ensure that the prediction was made before 12 pm) photovoltaic station data; firstly, converting the processed data set into three-dimensional data (samples, timeps, dimensions) by using a sliding window operation to adjust the data set to the input shape required by the proposed model; then screening the obtained data to ensure that the interval between two adjacent samples of the data is 24 hours, so that the corresponding photovoltaic power has a unique value, and model training, evaluation and visualization of rolling prediction are facilitated; and then completing construction of a day-ahead short-term prediction data form, wherein the obtained day-ahead prediction data form is shown in table 1.
TABLE 1 day-ahead prediction data
The implementation of the second aspect of the present application is that the invention model is constructed:
because the convolution kernel of the TCN model is fixed in size, in order to reduce the difficulty of extracting multi-scale space-time features from an input sequence, and in order to strengthen the information of the influence of the previous moment of extracting photovoltaic power on the current moment, a nonlinear mapping relation between multi-feature input and photovoltaic power is established.
As shown in FIG. 2, the system is a structure diagram of the MHATCN-TCN model, and comprises an input layer, an MHATCN-TCN module and an output layer;
wherein the input layer contains multi-feature data (any combination of photovoltaic output power (P) and total irradiance (Ti), normal direct irradiance (Ni), horizontal irradiance (Hi), air temperature (At), air pressure (Ap), relative humidity (Rh)), the multi-feature data comprising these seven terms, the letters being english abbreviations for only certain features in the invention; the multi-feature data is output to a first channel of the MHATCN-TCN module; a separate photovoltaic output power (P) data output channel second channel; i1 and I2 are the two-channel inputs of the MHATCN-TCN model respectively;
wherein the MHATCN-TCN module comprises a first channel and a second channel,
the processed multi-feature data is input into a first channel, and specific data is shown in fig. 3, wherein the first channel comprises a multi-head attention Mechanism (MHA) and a time convolutional neural network (TCN); the number of heads of the multi-head attention Mechanism (MHA) is the number of input features, different spatial mapping is performed on Q, K, V in different heads inside the multi-head attention Mechanism (MHA), and based on the design, each head may focus on different parts of the input, and may represent a more complex function than a simple weighted average; q, K, V of the multi-headed attentiveness Mechanism (MHA) uses the same input, its output shape is the same as the input shape by default, which facilitates the subsequent input time convolutional neural network (TCN); the output of the multi-headed attentiveness Mechanism (MHA) is then taken as an input to a time convolutional neural network (TCN) which may be used to summarize, extract, learn information entered by the multi-headed attentiveness Mechanism (MHA).
Specifically, the first-channel multi-head attention Mechanism (MHA)
Each attention head h i The computational expression of (i=1, …, h) is as follows:
where q (is Queries), k (is Keys), v (is Values) are given inputs,
for the trainable parameters, the attention pooling function f selects a scaled dot product attention.
Then the multi-head attention Mechanism (MHA) is output to a time convolution neural network (TCN) through another linear conversion, the TCN corresponds to the h head spliced results, and the trainable parameters are as follows
The expression is as follows:
the time convolutional neural network (TCN) uses a jump connection, followed by the time convolutional neural network (TCN) expression as follows:
wherein b is batch size, f 1 For the number of convolution kernels (filters) of a time convolutional neural network (TCN), F 1 The process is calculated for the existing TCN network.
The processed photovoltaic power is input into a second channel, and specific data are shown in fig. 4; the second channel adopts a double-layer TCN network; the double-layer TCN network consists of causal convolution, expansion convolution and residual error connection; the double-layer TCN not only has the advantages of parallelism and time sequence causality, but also can flexibly adjust the receptive field, so that the double-layer TCN is very suitable for processing time sequence data, and the time sequence relativity of the photovoltaic output power can be fully mined; specifically, the dual layer TCN network uses a hop connection, the input of the dual layer TCN network being
The expression is as follows:
wherein f 2 For the number of convolution kernels of the double-layer TCN network, F 2 The process is calculated for the existing TCN network.
The output layer comprises a full-connection neural network (Linear), and finally, the double-channel results are spliced in columns and input into the full-connection neural network, so that a mapping relation between multi-scale input and photovoltaic power is established; specifically, the time convolution neural network (TCN) filters of the two channels are set as f 1 Double-layer TCN network filters are set to f 2 The output is (b, f) 1 ) And (b, f) 2 ) After column splicing, (b, f), f is f 1 +f 2 The original data characteristics can be extracted and enriched by performing column splicing, and the learning capacity of the follow-up fully-connected neural network is enhanced; then, inputting into a full-connected neural network (Linear), using 3 full-connected layers, wherein the Leakyrlu is used as an activation function), and dropout prevents the model from being over-fitted; the full connection can effectively retain the extracted information, and establish a mapping relation between the multi-scale input and the photovoltaic power to obtain a photovoltaic power prediction sequence. Through tests, the model provided by the invention has better photovoltaic power prediction performance in the future and in the short term
Specifically, the calculation expression of the last full-join layer, i.e., the final output sequence Y, is as follows:
wherein s is the number of the set output sequences, C is the spliced result of the columns T1 and T2,
f＝f 1 +f 2 ，
and->
Is a trainable parameter.
The third aspect is implemented by performing optimal feature screening on multi-feature data input into the MHATCN-TCN model:
firstly, the proposed model uses different feature combinations to carry out experiments, the feature combination with the highest accuracy is obtained by trial, and the number of the features in the data set is seven, namely: photovoltaic output power (P), total irradiance (Ti), normal direct irradiance (Ni), horizontal irradiance (Hi), air temperature (At), air pressure (Ap), relative humidity (Rh). The parameter settings of the MHATCN-TCN model when performing the optimal feature screening are given in table 2.
TABLE 2 parameter settings of MHATCN-TCN model during optimal feature screening
According to the above parameter settings, and in order to represent the general level of each feature combination, the study used seven features in the dataset to perform 5 experiments on each of the different feature combinations and averaged, the experimental results are shown in table 3.
TABLE 3 test results of optimal feature screening
From the above results, photovoltaic output power (P), total irradiance (Ti), normal direct irradiance (Ni), horizontal scattered irradiance (Hi), air pressure (Ap) are the best feature combinations.
The fourth aspect is specifically implemented as evaluating the performance of the trained model;
next, the invention evaluates the performances of the MHATCN model and the MHATCN-TCN model from MAE, RMSE, R respectively 2 Judging experimental results at three angles; parameters of the MHATCN model and the MHATCN-TCN model are respectively set, and the specific parameter setting is shown in table 4;
TABLE 4 specific parameter settings for MHATCN model and MHATCN-TCN model
The evaluation index errors of the MHATCN-TCN model of the invention are MAE=1.115, RMSE=2.399 and R respectively 2 =0.880. According to the accuracy calculation formula, the accuracy of the MHATCN-TCN model is 96.28%.
MHATCN model evaluation index errors are mae=1.279, rmse=2.585, r respectively 2 ＝0.839。
Wherein, wherein: p (P) Mi For the actual power at time i, P Pi And (3) the predicted value of the daily short-term power at the moment i, cap is the capacity of the total assembly machine of the photovoltaic power station, and n is the number of samples.
The invention increases the receptive field by changing the convolution kernel size in the first channel and increases the receptive field by changing the convolution kernel size and the number of layers of the time convolution neural network in the second channel, so that the proposed model can consider the influence of the whole input sequence length (three and a half days) on the current output power.
Specifically, the first channel convolution Kernel size in the MHATCN-TCN model is set to kernel_size=29; the convolution Kernel size of the second channel is set kernel_size=15, nb_stacks=2; the reason for this arrangement enlarges the receptive field of the TCN layer, so that the TCN networks of the two channels both cover the whole input time step, so as to achieve the effect of completely extracting the time-space characteristic information of the whole time step, and as the nb_stacks of the first channel and the second channel are different, the convolution kernel size is nearly doubled according to the receptive field calculation formula, and the following calculation process of the receptive field is given:
R field ＝1+2·(K size -1)·Nb_stacks·∑d i
where Σdi represents the number of stacked expansion convolution layers, K, in each residual block size wei is the convolution kernel size, d is the coefficient of expansion, and nb_stacks is the number of residual blocks.
In practical application, the size of d can be set according to the time dimension length of the input data so that the receptive field can cover all input information; while the accuracy of the MHATCN model is drastically reduced when the MHATCN model is increased to the same receptive field, and the MAE is 1.638; therefore, the convolution kernel of the MHATCN model is adjusted to be small in size, and the difficulty of simultaneously processing a large amount of information by a single channel is reduced;
as shown in fig. 5 and fig. 6, the MHATCN model and the MHATCN-TCN model are respectively used for testing a part of the prediction result and a prediction power scatter diagram in the centralized test;
the prediction curve of fig. 5 and the scatter diagram of fig. 6 can show that the prediction precision of the MHATCN-TCN model is higher, and the fitting effect for the true value curve is better; the result shown in the graph is 3 days, and the MHATCN-TCN model has better prediction effect because the MHATCN-TCN model is provided with a channel for independently extracting photovoltaic output power information on the basis of the MHATCN model, then the channel is spliced with the MHATCN channel in a column manner to strengthen the original characteristics, finally the channel is connected with a fully connected neural network, and the factor influencing the photovoltaic output power is excavated by using the enhanced characteristics, so that a nonlinear relation is established; namely, the MHATCN-TCN specific parallel processing mode has stronger extraction and learning capacity on photovoltaic data of short period before the day and photovoltaic power of 1 day in the future.
Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This application is intended to cover any adaptations, uses, or adaptations of the disclosure following, in general, the principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains; it is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.
The embodiments of the present invention described above do not limit the scope of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of the present invention as set forth in the appended claims.
Claims (10)
1. A method for training a short-term photovoltaic power model before date, the method comprising:
after the multi-characteristic photovoltaic data are subjected to data preprocessing, the processed data set is constructed in a future prediction data set form based on time scale division;
inputting the data set after the division construction into a preset MHATCN-TCN model for training.
2. The method for training a short-term photovoltaic power model before day of claim 1, wherein: the time scale division is specifically based on that the data start-stop time is 3.5 days (336 points) and the label start-stop time is 1 day (96 points).
3. The method for training a short-term photovoltaic power model before day of claim 1, wherein: the data preprocessing is specifically to perform maximum and minimum normalization processing on data after processing of a null value and an abnormal value.
4. A method of training a short-term photovoltaic power model over the day according to claim 3, characterized by: the construction method of the form of the day-ahead prediction data set comprises the following steps: firstly, converting the processed data set into three-dimensional data (samples, timeps, dimensions) by using a sliding window operation to adjust the data set to the input shape required by the proposed model; the resulting data was then screened such that two adjacent samples of data were 24 hours apart.
5. The method for training a short-term photovoltaic power model before day of claim 1, wherein: the MHATCN-TCN model comprises an input layer, an MHATCN-TCN module and an output layer;
the input layer comprises multi-feature data, and all the multi-feature data are output to a first channel of the HATCN-TCN module; a separate photovoltaic output power (P) data output channel second channel;
the MHATCN-TCN module comprises a first channel and a second channel, wherein the first channel comprises a multi-head attention Mechanism (MHA) and a time convolutional neural network (TCN); the number of heads of a multi-head attention Mechanism (MHA) is the number of input features, and Q, K, V in different heads can be subjected to different spatial mapping in the multi-head attention mechanism; subsequently, taking the output of the multi-headed attentiveness Mechanism (MHA) as the input of a time convolutional neural network (TCN); the second channel adopts a double-layer TCN network; the double-layer TCN network consists of causal convolution, expansion convolution and residual error connection;
and finally, performing column splicing on the double-channel result, inputting the double-channel result into the full-connection neural network, and establishing a mapping relation between the multi-scale input and the photovoltaic power.
6. The method for training a short-term photovoltaic power model before day of claim 5, wherein: the first channel convolution Kernel size is set to kernel_size=29; setting kernel_size=15 for the convolution Kernel size of the second channel; nb_stacks=2.
7. The method for training a short-term photovoltaic power model before day of claim 5, wherein: the multi-feature data includes photovoltaic output power (P), total irradiance (Ti), normal direct irradiance (Ni), horizontal scattered irradiance (Hi), air pressure (Ap).
8. A solar short-term photovoltaic power prediction method is characterized by comprising the following steps of:
acquiring photovoltaic data of multiple characteristics to be tested;
inputting the photovoltaic data with multiple characteristics to be tested into the MHATCN-TCN model obtained by the training method of the short-term photovoltaic power model before day according to claim 1 to obtain the photovoltaic predicted power.
9. An electronic device comprising a memory and a processor, the memory being configured to store a computer program, the processor being configured to implement the method of training the short-term day photovoltaic power model of claim 1 or the method of predicting short-term day photovoltaic power of claim 8 when the computer program is executed.
10. A computer readable storage medium, wherein a computer program is stored on the storage medium, which when executed by a processor, implements the method of training a short-term day photovoltaic power model according to claim 1, or implements the short-term day photovoltaic power prediction method according to claim 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211550852.9A CN116090604A (en) | 2022-12-05 | 2022-12-05 | Training method, prediction method and device for photovoltaic power model in future and short term |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211550852.9A CN116090604A (en) | 2022-12-05 | 2022-12-05 | Training method, prediction method and device for photovoltaic power model in future and short term |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116090604A true CN116090604A (en) | 2023-05-09 |
Family
ID=86201557
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211550852.9A Pending CN116090604A (en) | 2022-12-05 | 2022-12-05 | Training method, prediction method and device for photovoltaic power model in future and short term |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116090604A (en) |
-
2022
- 2022-12-05 CN CN202211550852.9A patent/CN116090604A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108876054B (en) | Short-term power load prediction method based on improved genetic algorithm optimization extreme learning machine | |
| CN109597967B (en) | Power system distribution substation load data anomaly detection and repair method | |
| CN105303262A (en) | Short period load prediction method based on kernel principle component analysis and random forest | |
| CN112101480A (en) | Multivariate clustering and fused time sequence combined prediction method | |
| CN114239718B (en) | High-precision long-term time sequence prediction method based on multi-element time sequence data analysis | |
| CN116128141B (en) | Storm surge prediction method and device, storage medium and electronic equipment | |
| CN110726898A (en) | Power distribution network fault type identification method | |
| CN115688993A (en) | Short-term power load prediction method suitable for power distribution station area | |
| CN110991737A (en) | Ultra-short-term wind power prediction method based on deep belief network | |
| CN114330934A (en) | Model parameter self-adaptive GRU new energy short-term power generation power prediction method | |
| CN115238854A (en) | Short-term load prediction method based on TCN-LSTM-AM | |
| CN116599050A (en) | Photovoltaic prediction method and related device based on self-attention mechanism | |
| CN115375014A (en) | Source-load combination probability prediction method, device and storage medium | |
| CN115146739A (en) | Power transformer fault diagnosis method based on stacked time series network | |
| CN114118401A (en) | Neural network-based power distribution network flow prediction method, system, device and storage medium | |
| CN117668493A (en) | Tobacco equipment fault prediction method and system | |
| CN113627685A (en) | Wind power generator power prediction method considering wind power on-grid load limitation | |
| CN113837434A (en) | Solar photovoltaic power generation prediction method and device, electronic equipment and storage medium | |
| CN116629406A (en) | Power consumption carbon emission prediction method based on improved Prophet model | |
| CN117113202A (en) | Power loop energy consumption detection method and equipment based on joint error stacking model | |
| CN116885699A (en) | Power load prediction method based on dual-attention mechanism | |
| CN116090604A (en) | Training method, prediction method and device for photovoltaic power model in future and short term | |
| CN115759343A (en) | E-LSTM-based user electric quantity prediction method and device | |
| Ji et al. | Multi-indicators prediction in microservice using Granger causality test and Attention LSTM | |
| CN115545319A (en) | Power grid short-term load prediction method based on meteorological similar day set |
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 |