WO2016210102A1 - Method of forecasting for solar-based power systems - Google Patents
Method of forecasting for solar-based power systems Download PDFInfo
- Publication number
- WO2016210102A1 WO2016210102A1 PCT/US2016/038977 US2016038977W WO2016210102A1 WO 2016210102 A1 WO2016210102 A1 WO 2016210102A1 US 2016038977 W US2016038977 W US 2016038977W WO 2016210102 A1 WO2016210102 A1 WO 2016210102A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- forecasting
- solar
- time series
- based power
- power systems
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 64
- 230000002688 persistence Effects 0.000 claims description 59
- 238000012549 training Methods 0.000 claims description 39
- 238000013277 forecasting method Methods 0.000 claims description 22
- 238000005259 measurement Methods 0.000 claims description 18
- 239000013598 vector Substances 0.000 claims description 16
- 238000001914 filtration Methods 0.000 claims description 5
- 238000000638 solvent extraction Methods 0.000 claims description 5
- 238000010801 machine learning Methods 0.000 abstract description 10
- 238000013459 approach Methods 0.000 abstract description 8
- 230000002123 temporal effect Effects 0.000 description 20
- 230000005855 radiation Effects 0.000 description 13
- 238000011156 evaluation Methods 0.000 description 10
- 230000006870 function Effects 0.000 description 9
- 238000003066 decision tree Methods 0.000 description 7
- 238000012360 testing method Methods 0.000 description 7
- 238000004422 calculation algorithm Methods 0.000 description 6
- 230000036961 partial effect Effects 0.000 description 6
- 238000013145 classification model Methods 0.000 description 4
- 238000005266 casting Methods 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000007635 classification algorithm Methods 0.000 description 2
- 238000012937 correction Methods 0.000 description 2
- 238000012417 linear regression Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000013179 statistical model Methods 0.000 description 2
- 238000004577 artificial photosynthesis Methods 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000005094 computer simulation Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000000428 dust Substances 0.000 description 1
- 238000005183 dynamical system Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000010438 heat treatment Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 230000003449 preventive effect Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000001902 propagating effect Effects 0.000 description 1
- 238000013138 pruning Methods 0.000 description 1
- 238000003908 quality control method Methods 0.000 description 1
- 238000007637 random forest analysis Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 230000001932 seasonal effect Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 230000036962 time dependent Effects 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000009423 ventilation Methods 0.000 description 1
Classifications
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02S—GENERATION OF ELECTRIC POWER BY CONVERSION OF INFRARED RADIATION, VISIBLE LIGHT OR ULTRAVIOLET LIGHT, e.g. USING PHOTOVOLTAIC [PV] MODULES
- H02S50/00—Monitoring or testing of PV systems, e.g. load balancing or fault identification
- H02S50/10—Testing of PV devices, e.g. of PV modules or single PV cells
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01J—MEASUREMENT OF INTENSITY, VELOCITY, SPECTRAL CONTENT, POLARISATION, PHASE OR PULSE CHARACTERISTICS OF INFRARED, VISIBLE OR ULTRAVIOLET LIGHT; COLORIMETRY; RADIATION PYROMETRY
- G01J1/00—Photometry, e.g. photographic exposure meter
- G01J1/42—Photometry, e.g. photographic exposure meter using electric radiation detectors
- G01J1/4204—Photometry, e.g. photographic exposure meter using electric radiation detectors with determination of ambient light
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R21/00—Arrangements for measuring electric power or power factor
- G01R21/133—Arrangements for measuring electric power or power factor by using digital technique
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J3/00—Circuit arrangements for ac mains or ac distribution networks
- H02J3/38—Arrangements for parallely feeding a single network by two or more generators, converters or transformers
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02S—GENERATION OF ELECTRIC POWER BY CONVERSION OF INFRARED RADIATION, VISIBLE LIGHT OR ULTRAVIOLET LIGHT, e.g. USING PHOTOVOLTAIC [PV] MODULES
- H02S50/00—Monitoring or testing of PV systems, e.g. load balancing or fault identification
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01J—MEASUREMENT OF INTENSITY, VELOCITY, SPECTRAL CONTENT, POLARISATION, PHASE OR PULSE CHARACTERISTICS OF INFRARED, VISIBLE OR ULTRAVIOLET LIGHT; COLORIMETRY; RADIATION PYROMETRY
- G01J1/00—Photometry, e.g. photographic exposure meter
- G01J1/42—Photometry, e.g. photographic exposure meter using electric radiation detectors
- G01J2001/4266—Photometry, e.g. photographic exposure meter using electric radiation detectors for measuring solar light
- G01J2001/4276—Solar energy integrator over time
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01W—METEOROLOGY
- G01W1/00—Meteorology
- G01W1/10—Devices for predicting weather conditions
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
- H02J2203/20—Simulating, e g planning, reliability check, modelling or computer assisted design [CAD]
-
- 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
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/50—Photovoltaic [PV] energy
- Y02E10/56—Power conversion systems, e.g. maximum power point trackers
Definitions
- the present invention relates to a method of forecasting for solar-based power systems, and particularly a method for forecasting solar irradiance applied to photovoltaic systems and the like.
- the ability to forecast solar irradiance in near-real time is useful in managing power grid integration of renewable energy harnessed through such technologies as solar heating, photovoltaics (PV), solar thermal energy, solar architecture, and artificial photosynthesis.
- Solar irradiance is subject to sudden variations due to meteorological change, such as clouds, haze, and dust storms.
- meteorological change such as clouds, haze, and dust storms.
- sudden changes in solar irradiance can trigger grid instability.
- a cloud formation passing over a PV array can block up to 80% of the total irradiance reaching the PV array. Such a blockage would cause a rapid and steep fall in the power harnessed by the PV array, leading to unacceptable voltage deviations.
- Solar forecasting can help solve this problem by providing insights about forthcoming changes in solar irradiance that can be used to take preventive or reactive actions.
- Such actions may include the use of other energy sources to make up for the shortage of solar energy, a reduction of the rate of solar energy conversion to accommodate solar energy surplus, or storage of solar energy surplus.
- NPW Numerical Weather Prediction
- the statistical approach uses data mining techniques to train computational models on historical solar irradiation data, sometimes in conjunction with other meteorological data.
- NPW models provide a useful method to forecast solar irradiance beyond six hours and up to several days ahead, but are not appropriate for higher-resolution time forecasts (e.g., minutes) due to their coarse resolution.
- the spatial and temporal granularity of even the highest resolution NWP models such as the North American Mesoscale Forecast System (NAM), are insufficient to resolve most clouds and any patterns with characteristic timescales less than one hour.
- NAM North American Mesoscale Forecast System
- predictions based on radiometric measurements can be used in combination with statistical models to obtain predictions with a temporal step from 1 second to 15 minutes and a horizon of less than 1 hour;
- predictions based on sky cameras can be used to obtain predictions with a temporal step from 5 minutes to 15 minutes and a temporal horizon of less than 2-4 hours, depending on the location;
- satellite images can provide predictions of solar radiation with a temporal step from 5 minutes to 30 minutes, depending on the geostationary satellite used, and a temporal horizon of less than 6 hour.
- NWP, satellite and sky camera imaging techniques lack the spatial and temporal resolution to provide information regarding high temporal frequency fluctuations of solar irradiance.
- An alternative is provided through ground measurements of local meteorological conditions for temporal steps beyond 1 to 15 minutes.
- ARMA Autoregressive Moving Average
- ARIMA Autoregressive Integrated Moving Average
- CARDS Coupled Autoregressive and Dynamical System
- ANN Artificial Neural Network
- SVR Support Vector Regression
- the method of forecasting for solar-based power systems recognizes that no single solar irradiance forecasting model provides the best forecasting prediction for every current weather trend at every time of the year. Instead, the present method trains a classifier to select the best solar irradiance forecasting model for prevailing conditions through a machine learning approach. The resulting solar irradiance forecast predictions are then used to allocate the solar-based power systems resources and modify demand when necessary in order to maintain a substantially constant voltage supply in the system.
- the method of forecasting for solar-based power systems includes the following steps: (a) measuring solar irradiance parameters with sensors for a defined geographical region over predetermined time intervals to form a data set; (b) selecting a window size defining a number of past measurements and future forecast predictions to be made from the number of past measurements; (c) partitioning the data set into successive and adjacent time series training data sequences of the selected window size; (d) applying each of a selected plurality of forecasting methods to the time series training data sequences to obtain future forecast predictions from the forecasting methods applied; (e) comparing the future forecast predictions of each of the forecasting methods to measured data to obtain a corresponding error rate associated with each of the methods, given the time series training data sequences; (f) assigning the forecasting method with the lowest error rate as the forecasting class for the time series training data sequences; (g) repeating steps (a) through (f) to train a classifier to determine an optimal forecasting class for different time series training data sequences; (h) using the sensors to measure current solar
- the forecasting methods may be used as the forecasting methods, such as the aforementioned statistical and machine learning techniques.
- the Persistence technique and the Support Vector Regression (SVR) technique are each included in the selected forecasting methods, as well as various autoregressive (AR) models for the prediction of solar radiation in the short term (i.e., "now-casting") using ground measurements, such as radiometric measurements.
- SVR Support Vector Regression
- AR autoregressive
- This allows for temporal steps of predictions of 1, 5 and 10 minutes.
- the temporal horizon in all three cases is 15 temporal steps ahead.
- Fig. 1 is a block diagram illustrating system components for implementing a method of forecasting for solar-based power systems according to the present invention.
- Fig. 2A is a graph illustrating autocorrelation of a one-minute time series used in an embodiment of the method of forecasting for solar-based power systems.
- Fig. 2B is a graph illustrating partial autocorrelation of the one-minute time series used in the embodiment of the method used in Fig. 2A.
- Fig. 3A is a graph illustrating autocorrelation of a five-minute time series used in the embodiment of the method used in Fig. 2A.
- Fig. 3B is a graph illustrating partial autocorrelation of the five-minute time series used in the embodiment of the method used in Fig. 2A.
- Fig. 4A is a graph illustrating autocorrelation of a ten-minute time series used in the embodiment of the method used in Fig. 2A.
- Fig. 4B is a graph illustrating partial autocorrelation of the ten-minute time series used in the embodiment of the method used in Fig. 2A.
- Fig. 5 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to one-minute average time series data, with the data set being recorded in January of 2014.
- RMSD relative root mean squared deviation
- AR autoregressive
- PER persistence
- Fig. 6 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to one-minute average time series data, with a data set recorded in April of 2014.
- RMSD relative root mean squared deviation
- AR autoregressive
- PER persistence
- Fig. 7 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to one-minute average time series data, with a data set recorded in June of 2014.
- Fig. 8 is a graph comparing relative root mean squared deviation (RMSD, %) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for the one-minute time series.
- Fig. 9 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to five-minute average time series data, with the data set being recorded in January of 2014.
- RMSD relative root mean squared deviation
- AR autoregressive
- PER persistence
- Fig. 10 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to five-minute average time series data, with a data set recorded in June of 2014.
- RMSD relative root mean squared deviation
- AR autoregressive
- PER persistence
- Fig. 11 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to five-minute average time series data, with a data set recorded in November of 2014.
- RMSD relative root mean squared deviation
- AR autoregressive
- PER persistence
- Fig. 12 is a graph comparing relative root mean squared deviation (RMSD, %) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for the five-minute time series.
- RMSD relative root mean squared deviation
- Fig. 13 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to ten-minute average time series data, with the data set being recorded in February of 2014.
- RMSD relative root mean squared deviation
- AR autoregressive
- PER persistence
- Fig. 14 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to ten-minute average time series data, with a data set recorded in August of 2014.
- RMSD relative root mean squared deviation
- AR autoregressive
- PER persistence
- RMSD relative root mean squared deviation
- AR autoregressive
- PER persistence
- Fig. 16 is a graph comparing relative root mean squared deviation (RMSD, %) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for the ten-minute time series.
- RMSD relative root mean squared deviation
- the method of forecasting for solar-based power systems recognizes that no single solar irradiance forecasting model provides the best forecasting prediction for every current weather trend at every time of the year. Instead, the present method uses model evaluation data to train a classifier that enables the selection of the best solar irradiance forecasting model for prevailing conditions through a machine learning approach. The resulting solar irradiance forecast predictions are then used to allocate the solar-based power systems resources and modify demand when necessary in order to maintain a substantially constant voltage supply in the system.
- the method of forecasting for solar-based power systems includes the following steps: (a) measuring solar irradiance parameters with sensors for a defined geographical region over predetermined time intervals to form a data set; (b) selecting a window size defining a number of past measurements and future forecast predictions to be made from the number of past measurements; (c) partitioning the data set into successive and adjacent time series training data sequences of the selected window size; (d) applying each of a selected plurality of forecasting methods to the time series training data sequences to obtain future forecast predictions from the forecasting methods applied; (e) comparing the future forecast predictions of each of the forecasting methods to measured data to obtain a corresponding error rate associated with each of the methods, given the time series training data sequences; (f) assigning the forecasting method with the lowest error rate as the forecasting class for the time series training data sequences; (g) repeating steps (a) through (f) for the selected plurality of forecasting methods and use the resulting forecasting evaluation data to train a classifier to determine an optimal forecasting class
- step (a) above ground-measured solar radiation data is used to create data files, such as the exemplary data file shown below in Table 1.
- data files such as the exemplary data file shown below in Table 1.
- DNI direct normal irradiance
- GHI global horizontal irradiance
- DHI diffuse horizontal irradiance
- Kt and Kt _ > are clearness indices, calculated from the ground-measured GHI.
- Kt is the ratio of GHI to the calculated global horizontal radiation at the top of the atmosphere.
- Kt removes the seasonal dependence of GHI throughout the year.
- Kt _p is a modified form of Kt, which adds a correction factor for the changing atmospheric air masses traversed by the solar radiation at any moment.
- Kt and Kt _p are given by:
- Data filtering may be applied to the data set generated in step (a). Any suitable technique, such as normalization, discretization, or wavelet analysis, used alone or in combination, may be used to remove noise from the data. With normalization, all numeric values in the given data set are mapped into a standard numerical interval, typically 0-1. Discretization maps continuous numeric values into discrete counterparts (e.g., high, medium, low). The wavelet transform enables the decomposition of a time series into a time dependent sum of frequency components.
- One or several forecasting algorithms are then applied to T and the results are evaluated using an evaluation measure such as the Coefficient of Variation of the Root Mean Square Error (cvRMSE) to compare predicted (pred) against observed (obs) values.
- cvRMSE Coefficient of Variation of the Root Mean Square Error
- the window size, i is successively increased and decreased until the best evaluation results (e.g., the lowest cvRMSE) are found, to determine the best training window size j.
- step (c) the training data is partitioned into successive and adjacent time series training data sequences of the selected window size. For example, if the time scale is minutes, an ideal training window size may be 180 minutes, and the time series data for Kt _p would be as indicated in Table 2 below.
- each of a plurality of forecasting methods is applied to the time series training data sequences to obtain future forecast predictions from the forecasting method applied.
- the different forecasting models are developed to predict n-steps ahead for the solar irradiance variable of interest, using as training data the time series data developed above.
- two forecasting models are exemplified: Persistence and Support Vector Regression (SVR).
- SVR Persistence and Support Vector Regression
- the persistence model assumes that no change occurs from the present state throughout the forecasting period.
- each training sample is a pair ⁇ x, y ⁇ , where x £ W 1 is a vector for the time- series class to be learned, and y £ K y is the associated value.
- the aim of the machine learning algorithm is to find a function such that each x[ in the training dataset approximates its value y as closely as possible.
- the resulting function is then used to predict values n- steps ahead of the time series data used for training.
- the objective of regression is to learn the weight vector w that has the smallest possible length so as to avoid over-fitting.
- a given margin of deviation ⁇ is allowed with no penalty, and a given margin ⁇ is specified where deviation is allowed with increasing penalty.
- the length of the weight vector w is obtained by minimizing the loss function, ⁇
- 2 + C ⁇ 1 ( ⁇ ; + ⁇ * ) 1 ' subject to the constraints y ⁇ — (w ⁇ xl + b) ⁇ ⁇ + ⁇ , or y — (w ⁇ x[ + b) ⁇ ⁇ — ⁇ * , for ⁇ , ⁇ * ⁇ 0.
- the solution is given by Yi * )(w ⁇ x ) + b, where a and a * are Lagrange multipliers.
- the training vectors giving nonzero Lagrange multipliers are called support vectors and are used to construct the regression function. If the input data are not amenable to linear regression, then the vector data are mapped into a higher dimensional features space using a kernel function ⁇ .
- Forecasting models are next evaluated, using an evaluation measure such as cvRMSE. Forecasting evaluation results relative to each prediction step are stored for each training data sequence (e.g., the 180 minutes used above), as shown in Table 3.
- the forecasting models are then saved as software components that take as input a time series data sequence to output predictions for m steps ahead.
- classification models are developed to selected the best forecasting method.
- a time series data set is created that includes the evaluated forecasting models and their time series training data, as shown below in Table 4.
- Table 4 Time Series Data Created to Train Classification Models
- step (f) the forecasting method with the lowest error rate is assigned as the forecasting class for the time series training data sequences.
- the best performing class is established as the model with the best evaluation results; e.g., a lower cvRMSE, as shown
- Additional data filtering processes may be applied to the above data set.
- Data can be filtered using any suitable filtering techniques, such as discretization, normalization or wavelet transforms, or a combination of these.
- the data is then used to train a classification model capable of recognizing the highest ranking model given an input time series training data sequence.
- a number of machine learning algorithms can be used to train a classifier from data, such as, for example, the decision-tree classification algorithm.
- decision-tree classification the model identifies members of a class as the result of a sequence of decisions.
- a decision tree typically consists of two types of nodes: test nodes and prediction nodes. The test node describes the condition that must be met in order to make a decision. Several test nodes can occur in a sequence to indicate the number of decisions that must be taken and the order in which these decisions follow one another to reach a prediction outcome.
- a decision tree classifier is "learned" from a training dataset by a model which establishes the sequential order of test nodes according to how informative the nodes' attributes are.
- the model determines the information content of an attribute by its information gain with respect to the classification tasks.
- the information gain of an attribute with respect to a class is the reduction in entropy (i.e., the uncertainty) of the value for the class the value of the attribute is known.
- the test nodes with more informative attributes occur earlier in the decision tree.
- the model creates test nodes using the available attributes until all data in the training dataset have been accounted for. Typically, not all attributes are used because decision tree learners use pruning strategies to reduce the number of nodes. The number of attributes depends on the specific implementation.
- the "alternating decision tree algorithm” uses a machine learning meta-algorithm, called “boosting”, to minimize the number of nodes without losing accuracy.
- boosting a machine learning meta-algorithm
- any suitable classification algorithm may be used, such as, for example, Bayesian nets, Support Vector Machines, boosting, Naive Bayes, bagging, random forest and Model Trees.
- the classification model is saved as a software component capable of recognizing the highest ranking model given an input time series training data sequence.
- Fig. 1 Data is entered into system 10 via any suitable type of user interface 16, and may be stored in memory 12, which may be any suitable type of computer readable and programmable memory and is preferably a non- transitory, computer readable storage medium. Calculations are performed by processor 14, which may be any suitable type of computer processor and may be displayed to the user on display 18, which may be any suitable type of computer display.
- Sensor data is collected from solar irradiance sensors 20, such as, for example, satellite and sky cameras, radiometric sensors, pyrheliometers, pyranometers and the like.
- Processor 14 may be associated with, or incorporated into, any suitable type of computing device, for example, a personal computer or a programmable logic controller.
- the display 18, the processor 14, the memory 12 and any associated computer readable recording media are in communication with one another by any suitable type of data bus, as is well known in the art.
- Examples of computer-readable recording media include non-transitory storage media, a magnetic recording apparatus, an optical disk, a magneto-optical disk, and/or a semiconductor memory (for example, RAM, ROM, etc.).
- Examples of magnetic recording apparatus that may be used in addition to memory 112, or in place of memory 12, include a hard disk device (HDD), a flexible disk (FD), and a magnetic tape (MT).
- Examples of the optical disk include a DVD (Digital Versatile Disc), a DVD-RAM, a CD-ROM (Compact Disc-Read Only Memory), and a CD-R (Recordable)/RW.
- non- transitory computer-readable storage media include all computer-readable media, with the sole exception being a transitory, propagating signal.
- the Persistence technique and the Support Vector Regression (SVR) technique may be used as the forecasting methods.
- the Persistence model may be replaced by autoregressive (AR) models, particularly for the prediction of solar radiation in the short term (i.e., "now-casting") using ground measurements, such as radiometric measurements. This allows for temporal steps of predictions of 1, 5 and 10 minutes. The temporal horizon in all three cases is 15 temporal steps ahead.
- AR autoregressive
- AR(/?) refers to the autoregressive model of order p.
- order p 1,..., 20 for 1, 5 and 10 minutes average time series.
- Exemplary data are from the Qatar Environment and Energy Research Institute (QEERI), which has been operating a high precision solar radiation monitoring station since the end of November 2012 in Education City, Doha (25.33°N, 51.43 ⁇ ).
- the station is equipped with a solar tracker with a sun sensor kit, for improved tracking accuracy, and a shading ball assembly for diffuse measurements.
- Mounted on the sun tracker are one first class pyrheliometer for measuring Direct Normal Irradiance (DNI), and two secondary standard pyranometers (one of them shaded) for Global Horizontal Irradiance (GHI) and Diffuse Horizontal Irradiance (DHI) measurements. Both pyranometers are fitted with ventilation units. Data from the monitoring station are sampled every second and recorded as minute averages in W/m 2 .
- DNI Direct Normal Irradiance
- GHI Global Horizontal Irradiance
- DHI Diffuse Horizontal Irradiance
- Figs. 2 A and 2B show autocorrelation and partial autocorrelation of a one-minute time series used in the alternative embodiment of the method of forecasting for solar-based power systems, respectively.
- Figs. 3A and 3B show autocorrelation and partial autocorrelation of a five-minute time series, respectively
- Figs. 4A and 4B show autocorrelation and partial autocorrelation of a ten- minute time series.
- the metrics to measure the errors rates of the models are the mean bias deviation (MBD), the relative root mean squared deviation (RMSD) and its relative values rMBD and rRMSD normalized with the mean value of the observed variable to predict for the period under validation.
- the present models are compared with a baseline model, with the intention of measuring the improvement achieved.
- the basic model chosen is the persistence model (PER) because it is the most extended model to contrast new proposed models.
- Fig. 5 is a graph comparing relative root mean squared deviation (RMSD) of the autocorrelation (AR) model against the persistence (PER) model, and a combination autocorrelation model in the alternative embodiment of the method of forecasting for solar- based power systems, specifically for each horizon of prediction applied to one-minute average time series data (taken from January of 2014).
- the new "AR combined" model (AR_COMB_MIN) is detailed below.
- Figs. 6 and 7 show similar comparisons for data sets recorded in April of 2014 and June of 2014, respectively.
- Fig. 8 shows model rRMSD results for the one-minute time series for each temporal horizon of prediction.
- the above data shows that the AR-combined model is better than any other model, including the PER baseline model. The difference is quite considerable during the entire year, except in June. The lower limit of monthly rRMSD is around 6%.
- Fig. 9 is a graph comparing relative root mean squared deviation (RMSD) of the autocorrelation (AR) model against the persistence (PER) model, and a combination autocorrelation model in the alternative embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to the five-minute average time series data (taken from January of 2014).
- Figs. 10 and 11 show similar comparisons for data sets recorded in June of 2014 and November of 2014, respectively.
- Table 7 shows MBD, RMSD, rMBD and rRMSD results for the AR, PER and AR- combined models with the five-minute time series from 2014.
- Fig. 12 shows model rRMSD results for the five-minute time series for each temporal horizon of prediction. From the above data, one can conclude that the AR-combined model outperforms the other models and the PER baseline model for the case of a five-minute average time series. The difference is quite considerable during the entire year, except in June. The lower limit of monthly rRMSD is around 6%.
- Fig. 13 is a graph comparing relative root mean squared deviation (RMSD) of the autocorrelation (AR) model against the persistence (PER) model, and the combination autocorrelation model in the alternative embodiment of the method of forecasting for solar- based power systems, specifically for each horizon of prediction applied to the ten- minute average time series data (taken from February of 2014).
- Figs. 14 and 15 show similar comparisons for data sets recorded in August of 2014 and December of 2014, respectively.
- Table 8 shows MBD, RMSD, rMBD and rRMSD results for the AR, PER and AR- combined models with the ten- minute time series from 2014.
- Fig. 16 shows model rRMSD results for the ten-minute time series for each temporal horizon of prediction. From the above data, one can conclude that the AR-combined model is better than any AR model and the PER model, also for the case of 10 minute average time series. The difference is quite considerable during the entire year, except in June. This is due to the fact that June is a month where there are almost no clouds in Kuwait and almost no impact in lower variability of solar radiation. The lower limit of monthly rRMSD is around 9%. It is to be understood that the present invention is not limited to the embodiments described above, but encompasses any and all embodiments within the scope of the following claims.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Power Engineering (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Sustainable Development (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The method of forecasting for solar-based power systems (10) recognizes that no single solar irradiance forecasting model provides the best forecasting prediction for every current weather trend at every time of the year. Instead, the method trains a classifier to select the best solar irradiance forecasting model for prevailing conditions through a machine learning approach. The resulting solar irradiance forecast predictions are then used to allocate the solar-based power systems (10) resources and modify demand when necessary in order to maintain a substantially constant voltage supply in the system (10).
Description
METHOD OF FORECASTING FOR SOLAR-BASED POWER SYSTEMS
TECHNICAL FIELD
The present invention relates to a method of forecasting for solar-based power systems, and particularly a method for forecasting solar irradiance applied to photovoltaic systems and the like.
BACKGROUND ART
The ability to forecast solar irradiance in near-real time is useful in managing power grid integration of renewable energy harnessed through such technologies as solar heating, photovoltaics (PV), solar thermal energy, solar architecture, and artificial photosynthesis. Solar irradiance is subject to sudden variations due to meteorological change, such as clouds, haze, and dust storms. When significant amounts of solar energy are introduced into the power grid, sudden changes in solar irradiance can trigger grid instability. For example, a cloud formation passing over a PV array can block up to 80% of the total irradiance reaching the PV array. Such a blockage would cause a rapid and steep fall in the power harnessed by the PV array, leading to unacceptable voltage deviations. Solar forecasting can help solve this problem by providing insights about forthcoming changes in solar irradiance that can be used to take preventive or reactive actions. Such actions may include the use of other energy sources to make up for the shortage of solar energy, a reduction of the rate of solar energy conversion to accommodate solar energy surplus, or storage of solar energy surplus.
Solar forecasting is typically carried out using physical or statistical approaches. The physical approach relies on Numerical Weather Prediction (NWP) models, which use mathematical models of the atmosphere and oceans to predict the evolution of the atmosphere from initial conditions. The statistical approach uses data mining techniques to train computational models on historical solar irradiation data, sometimes in conjunction with other meteorological data. NPW models provide a useful method to forecast solar irradiance beyond six hours and up to several days ahead, but are not appropriate for higher-resolution time forecasts (e.g., minutes) due to their coarse resolution. More specifically, the spatial and temporal granularity of even the highest resolution NWP models, such as the North American Mesoscale Forecast System (NAM), are insufficient to resolve most clouds and any patterns with characteristic timescales less than one hour.
Depending on the kind of instrument used in combination with statistical models, one can differentiate three kind of predictions, depending on the horizon of prediction: (1) predictions based on radiometric measurements can be used in combination with statistical models to obtain predictions with a temporal step from 1 second to 15 minutes and a horizon of less than 1 hour; (2) predictions based on sky cameras can be used to obtain predictions with a temporal step from 5 minutes to 15 minutes and a temporal horizon of less than 2-4 hours, depending on the location; and (3) satellite images can provide predictions of solar radiation with a temporal step from 5 minutes to 30 minutes, depending on the geostationary satellite used, and a temporal horizon of less than 6 hour. The predictions from these three instruments and models can be superposed, and sometimes one can use a combination of their predictions. Both NWP, satellite and sky camera imaging techniques lack the spatial and temporal resolution to provide information regarding high temporal frequency fluctuations of solar irradiance. An alternative is provided through ground measurements of local meteorological conditions for temporal steps beyond 1 to 15 minutes.
With higher-resolution timescales, machine learning and statistical techniques have been shown to provide an effective methodology for solar forecasting. Various statistical and machine learning techniques have been used to forecast solar irradiance, including
Autoregressive Moving Average (ARMA), Autoregressive Integrated Moving Average (ARIMA), Coupled Autoregressive and Dynamical System (CARDS), Artificial Neural Network (ANN), and Support Vector Regression (SVR). These algorithms have also been successfully improved through combination with data filtering techniques, such as wavelet transforms. For evaluation purposes, the Persistence model, according to which no difference is assumed between current and future irradiance values, is usually used as a baseline.
One of the main problems with solar now-casting is that no single forecasting model can consistently provide superior forecasts in all prediction instances. Evaluation of forecasting results against observed data show that different forecasting approaches, including persistence, can rival each other across non- aggregated (e.g., minute by minute) forecast units. Thus, a method of forecasting for solar-based power systems solving the aforementioned problems is desired.
DISCLOSURE OF INVENTION
The method of forecasting for solar-based power systems recognizes that no single solar irradiance forecasting model provides the best forecasting prediction for every current
weather trend at every time of the year. Instead, the present method trains a classifier to select the best solar irradiance forecasting model for prevailing conditions through a machine learning approach. The resulting solar irradiance forecast predictions are then used to allocate the solar-based power systems resources and modify demand when necessary in order to maintain a substantially constant voltage supply in the system.
The method of forecasting for solar-based power systems includes the following steps: (a) measuring solar irradiance parameters with sensors for a defined geographical region over predetermined time intervals to form a data set; (b) selecting a window size defining a number of past measurements and future forecast predictions to be made from the number of past measurements; (c) partitioning the data set into successive and adjacent time series training data sequences of the selected window size; (d) applying each of a selected plurality of forecasting methods to the time series training data sequences to obtain future forecast predictions from the forecasting methods applied; (e) comparing the future forecast predictions of each of the forecasting methods to measured data to obtain a corresponding error rate associated with each of the methods, given the time series training data sequences; (f) assigning the forecasting method with the lowest error rate as the forecasting class for the time series training data sequences; (g) repeating steps (a) through (f) to train a classifier to determine an optimal forecasting class for different time series training data sequences; (h) using the sensors to measure current solar irradiance parameters; (i) using the classifier to determine the optimal forecasting class for the current solar irradiance parameters; (j) making future forecast predictions from the current solar irradiance parameters using the optimal forecasting class; (k) predicting solar-based power system demands and generating capacities based upon the future forecast predictions made in step (j); and (1) making adjustments in the solar-based power system demands and stored energy in order to maintain a substantially constant voltage supply for the geographic region.
In the above method, several modeling techniques may be used as the forecasting methods, such as the aforementioned statistical and machine learning techniques. Preferably, the Persistence technique and the Support Vector Regression (SVR) technique are each included in the selected forecasting methods, as well as various autoregressive (AR) models for the prediction of solar radiation in the short term (i.e., "now-casting") using ground measurements, such as radiometric measurements. This allows for temporal steps of predictions of 1, 5 and 10 minutes. The temporal horizon in all three cases is 15 temporal steps ahead.
These and other features of the present invention will become readily apparent upon further review of the following specification and drawings.
BRIEF DESCRIPTION OF DRAWINGS
Fig. 1 is a block diagram illustrating system components for implementing a method of forecasting for solar-based power systems according to the present invention.
Fig. 2A is a graph illustrating autocorrelation of a one-minute time series used in an embodiment of the method of forecasting for solar-based power systems.
Fig. 2B is a graph illustrating partial autocorrelation of the one-minute time series used in the embodiment of the method used in Fig. 2A.
Fig. 3A is a graph illustrating autocorrelation of a five-minute time series used in the embodiment of the method used in Fig. 2A.
Fig. 3B is a graph illustrating partial autocorrelation of the five-minute time series used in the embodiment of the method used in Fig. 2A.
Fig. 4A is a graph illustrating autocorrelation of a ten-minute time series used in the embodiment of the method used in Fig. 2A.
Fig. 4B is a graph illustrating partial autocorrelation of the ten-minute time series used in the embodiment of the method used in Fig. 2A.
Fig. 5 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to one-minute average time series data, with the data set being recorded in January of 2014.
Fig. 6 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to one-minute average time series data, with a data set recorded in April of 2014.
Fig. 7 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to one-minute average time series data, with a data set recorded in June of 2014.
Fig. 8 is a graph comparing relative root mean squared deviation (RMSD, %) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for the one-minute time series.
Fig. 9 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to five-minute average time series data, with the data set being recorded in January of 2014.
Fig. 10 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to five-minute average time series data, with a data set recorded in June of 2014.
Fig. 11 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to five-minute average time series data, with a data set recorded in November of 2014.
Fig. 12 is a graph comparing relative root mean squared deviation (RMSD, %) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for the five-minute time series.
Fig. 13 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to ten-minute average time series data, with the data set being recorded in February of 2014.
Fig. 14 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to ten-minute average time series data, with a data set recorded in August of 2014.
Fig. 15 is a graph comparing relative root mean squared deviation (RMSD) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to ten-minute average time series data, with a data set recorded in December of 2014.
Fig. 16 is a graph comparing relative root mean squared deviation (RMSD, %) of the autoregressive (AR) models of order 1-20 against the persistence (PER) model, and a combination of the autoregressive models in an embodiment of the method of forecasting for solar-based power systems, specifically for the ten-minute time series.
Similar reference characters denote corresponding features consistently throughout the attached drawings.
BEST MODES FOR CARRYING OUT THE INVENTION
The method of forecasting for solar-based power systems recognizes that no single solar irradiance forecasting model provides the best forecasting prediction for every current weather trend at every time of the year. Instead, the present method uses model evaluation data to train a classifier that enables the selection of the best solar irradiance forecasting model for prevailing conditions through a machine learning approach. The resulting solar irradiance forecast predictions are then used to allocate the solar-based power systems resources and modify demand when necessary in order to maintain a substantially constant voltage supply in the system.
Generally, the method of forecasting for solar-based power systems includes the following steps: (a) measuring solar irradiance parameters with sensors for a defined geographical region over predetermined time intervals to form a data set; (b) selecting a window size defining a number of past measurements and future forecast predictions to be made from the number of past measurements; (c) partitioning the data set into successive and adjacent time series training data sequences of the selected window size; (d) applying each of a selected plurality of forecasting methods to the time series training data sequences to obtain future forecast predictions from the forecasting methods applied; (e) comparing the future forecast predictions of each of the forecasting methods to measured data to obtain a corresponding error rate associated with each of the methods, given the time series training data sequences; (f) assigning the forecasting method with the lowest error rate as the forecasting class for the time series training data sequences; (g) repeating steps (a) through (f)
for the selected plurality of forecasting methods and use the resulting forecasting evaluation data to train a classifier to determine an optimal forecasting class for different time series training data sequences; (h) using the sensors to measure current solar irradiance parameters; (i) using the classifier to determine the optimal forecasting class for the current solar irradiance parameters; (j) making future forecast predictions from the current solar irradiance parameters using the optimal forecasting class; (k) predicting solar-based power system demands and generating capacities based upon the future forecast predictions made in step (j); and (1) making adjustments in the solar-based power system demands and stored energy in order to maintain a substantially constant voltage supply for the geographic region.
In step (a) above, ground-measured solar radiation data is used to create data files, such as the exemplary data file shown below in Table 1. In Table 1 , direct normal irradiance (DNI), global horizontal irradiance (GHI) and diffuse horizontal irradiance (DHI) are each provided as one-minute averages of the measured solar radiation components (measured in W/m2). Each component in Table 1 was measured by a different sensor mounted on a high- precision solar radiation monitoring station.
Table 1 : Measured Solar Radiation Parameters
YYYY-MM-DD-HH-MM DNI GHI DHI Kt Kt _p
2014-03-05-06-47 305.34 167.55 102.75 0.65 0.92
2014-03-05-06-48 270.89 156.47 97.45 0.60 0.84
2014-03-05-06-49 221.88 141.53 91.46 0.53 0.74
2014-03-05-06-50 179.62 128.26 86.32 0.47 0.65
2014-03-05-06-51 84.08 105.34 82.82 0.38 0.52
2014-03-05-06-52 108.30 113.37 85.17 0.40 0.55
2014-03-05-09-46 624.76 688.97 225.32 0.681 0.703
In the above, Kt and Kt _ > are clearness indices, calculated from the ground-measured GHI. Kt is the ratio of GHI to the calculated global horizontal radiation at the top of the atmosphere. Kt removes the seasonal dependence of GHI throughout the year. Kt _p is a modified form of Kt, which adds a correction factor for the changing atmospheric air masses traversed by the solar radiation at any moment. Kt and Kt _p are given by:
GHI
Kt = (1)
GHItoa
Kt
Ktp— (2)
1.031e_1-4/(0-9+9-4/ m) + 0.1
1
am = (3)
cos(SZA) + 0.50573 (96.07995 - SZA)~1 6364 ' where GHItoa is the extraterrestrial solar radiation on a horizontal surface at the top of the atmosphere, am is the air mass, and SZA is the solar zenith angle in degrees. In order to ensure good accuracy in the modeling results, only measured data that pass some quality control tests should be included. For the data given above, the recommended Baseline Surface Radiation Network quality tests were applied to all entries. Several or just a single measure of solar irradiance can be used. In the present embodiment, all examples will be given with reference to the single measure Kt _p.
Data filtering may be applied to the data set generated in step (a). Any suitable technique, such as normalization, discretization, or wavelet analysis, used alone or in combination, may be used to remove noise from the data. With normalization, all numeric values in the given data set are mapped into a standard numerical interval, typically 0-1. Discretization maps continuous numeric values into discrete counterparts (e.g., high, medium, low). The wavelet transform enables the decomposition of a time series into a time dependent sum of frequency components.
In order to select the window size defining a number of past measurements and future forecast predictions to be made from the number of past measurements (step (b)), the number of steps ahead to be forecasted are first selected. Then, several data points are selected, at the beginning, middle and end of the data collection. For each data point, an initial training data set T is created with a window size, i, given by ί = n + m, where m is the number of steps ahead to be forecasted, and n is some function over m, e.g., 3m. One or several forecasting algorithms are then applied to T and the results are evaluated using an evaluation measure such as the Coefficient of Variation of the Root Mean Square Error (cvRMSE) to compare predicted (pred) against observed (obs) values. Here,
mean(obsti-n)
The window size, i, is successively increased and decreased until the best evaluation results (e.g., the lowest cvRMSE) are found, to determine the best training window size j.
In step (c), the training data is partitioned into successive and adjacent time series training data sequences of the selected window size. For example, if the time scale is
minutes, an ideal training window size may be 180 minutes, and the time series data for Kt _p would be as indicated in Table 2 below.
Table 2: Exemplary Partitioning of Training Data
In step (d), each of a plurality of forecasting methods is applied to the time series training data sequences to obtain future forecast predictions from the forecasting method applied. The different forecasting models are developed to predict n-steps ahead for the solar irradiance variable of interest, using as training data the time series data developed above. In the present embodiment, two forecasting models are exemplified: Persistence and Support Vector Regression (SVR). The persistence model assumes that no change occurs from the present state throughout the forecasting period. In SVR, within a machine learning approach to forecasting, each training sample is a pair {x, y}, where x £ W1 is a vector for the time- series class to be learned, and y £ K y is the associated value. The aim of the machine learning algorithm is to find a function such that each x[ in the training dataset approximates its value y as closely as possible. The resulting function is then used to predict values n- steps ahead of the time series data used for training. When the input data are amenable to linear regression, the SVR prediction function is given by y = w■ x[ + b, where ί = 1, ... , , w is the weight vector (i.e., a linear combination of training patterns that supports the regression function), x[ is the input vector (i.e., the training sample), y is the value for the input vector, and b is the bias (i.e., an average over marginal vectors, which are weight vectors that lie within the margins set by the loss function).
The objective of regression is to learn the weight vector w that has the smallest possible length so as to avoid over-fitting. To ease the regression task, a given margin of
deviation ε is allowed with no penalty, and a given margin ξ is specified where deviation is allowed with increasing penalty. The length of the weight vector w is obtained by minimizing the loss function, ^ ||w||2 + C∑=1(< ; + ^*)1' subject to the constraints y{— (w■ xl + b) < ε + ξι, or y — (w■ x[ + b)≥ ε— ξ*, for ξι, ξ*≥ 0. The solution is given by Yi
*)(w■ x ) + b, where a and a* are Lagrange multipliers. The training vectors giving nonzero Lagrange multipliers are called support vectors and are used to construct the regression function. If the input data are not amenable to linear regression, then the vector data are mapped into a higher dimensional features space using a kernel function Φ. One example is the polynomial kernel, according to which Φ(ϊν)■ Φ (¾ = (1 + w■ ¾3.
The forecasting models are next evaluated, using an evaluation measure such as cvRMSE. Forecasting evaluation results relative to each prediction step are stored for each training data sequence (e.g., the 180 minutes used above), as shown in Table 3.
Table 3: Forecasting Evaluation Results
cvRMSE cvRMSE cvRMSE cvRMSE
for for SVR for for SVR
Persistence
Period for Kt _ > Prediction model at Persistence ... model at
model at n
1 step model at 1 n steps
steps ahead step ahead ahead
ahead
2014-12- 01-06- -57→
1.941569 3.522824
2014- 12-01- 09-56
2014-12- 01-06- -57→
4.299455 8.545602
2014- 12-01- 09-56
2014-12- 01-06- -57→
7.165063 14.192361
2014- 12-01- 09-56
2014-12- 01-06- -57→
4.637791 14.184008
2014- 12-01- 09-56
2014-12- 01-06- -57→
5.841919 5.199473
2014- 12-01- 09-56
2014-12- 01-06- -57→ 25.46936
2014- 12-01- 09-56
The forecasting models are then saved as software components that take as input a time series data sequence to output predictions for m steps ahead. Next, classification models are developed to selected the best forecasting method. A time series data set is created that includes the evaluated forecasting models and their time series training data, as shown below in Table 4.
Table 4: Time Series Data Created to Train Classification Models
cvRMSE
cvRMSE
for
for SVR
Persistence
model Kt y at model Kt _jj at
Period for Kt _p Prediction (average minute
(average minute 1
over all 180 over all
steps
steps
ahead)
ahead)
2014-12- 01-06- ■57→
1.9416 3.5228 0.0801 0.2810
2014- 12-01- 09-56
2014-12- 01-06- ■57→
4.2995 8.5456 0.0801 0.2810
2014- 12-01- 09-56
2014-12- 01-06- ■57→
7.1651 14.1924 0.0801 0.2810
2014- 12-01- 09-56
2014-12- 01-06- ■57→
4.6378 14.1840 0.0801 0.2810
2014- 12-01- 09-56
2014-12- 01-06 ■57→
5.8419 5.1995 0.0801 0.2810
2014- 12-01- 09-56
2014-12- 01-06- ■57→
25.4694 15.1438 0.0801 0.2810
2014- 12-01- 09-56
In step (f), the forecasting method with the lowest error rate is assigned as the forecasting class for the time series training data sequences. The best performing class is established as the model with the best evaluation results; e.g., a lower cvRMSE, as shown
Table 5.
Table 5: Model Class Assignment to Training Records
cvRMSE
cvRMSE
for
for SVR
Persistence Kt _p model Kt _jj at Mode model at
Period for Kt _p Prediction (average minute 1
(average minute over all 1 Class over all 180 steps
steps
ahead)
ahead)
2014-12-01-06-57→
1.9416 3.5228 0.0801 . .. 0.2810 SVR 2014-12-01-09-56
2014-12-01-06-57→
4.2995 8.5456 0.0801 . .. 0.2810 SVR 2014-12-01-09-56
2014-12-01-06-57→
7.1651 14.1924 0.0801 . .. 0.2810 SVR 2014-12-01-09-56
2014-12-01-06-57→
4.6378 14.1840 0.0801 . .. 0.2810 SVR 2014-12-01-09-56
2014-12-01-06-57→
5.8419 5.1995 0.0801 . .. 0.2810 PER 2014-12-01-09-56
2014-12-01-06-57→
25.4694 15.1438 0.0801 0.2810 PER 2014-12-01-09-56
Additional data filtering processes may be applied to the above data set. Data can be filtered using any suitable filtering techniques, such as discretization, normalization or wavelet transforms, or a combination of these. The data is then used to train a classification model capable of recognizing the highest ranking model given an input time series training data sequence. A number of machine learning algorithms can be used to train a classifier from data, such as, for example, the decision-tree classification algorithm. In decision-tree classification, the model identifies members of a class as the result of a sequence of decisions. A decision tree typically consists of two types of nodes: test nodes and prediction nodes. The test node describes the condition that must be met in order to make a decision. Several test nodes can occur in a sequence to indicate the number of decisions that must be taken and the order in which these decisions follow one another to reach a prediction outcome.
A decision tree classifier is "learned" from a training dataset by a model which establishes the sequential order of test nodes according to how informative the nodes' attributes are. The model determines the information content of an attribute by its information gain with respect to the classification tasks. The information gain of an attribute with respect to a class is the reduction in entropy (i.e., the uncertainty) of the value for the class the value of the attribute is known. The test nodes with more informative attributes occur earlier in the decision tree. The model creates test nodes using the available attributes until all data in the training dataset have been accounted for. Typically, not all attributes are used because decision tree learners use pruning strategies to reduce the number of nodes. The number of attributes depends on the specific implementation. The "alternating decision tree algorithm" uses a machine learning meta-algorithm, called "boosting", to minimize the number of nodes without losing accuracy. It should be understood that any suitable classification algorithm may be used, such as, for example, Bayesian nets, Support Vector Machines, boosting, Naive Bayes, bagging, random forest and Model Trees. The classification model is saved as a software component capable of recognizing the highest ranking model given an input time series training data sequence.
It should be understood that the calculations may be performed by any suitable computer system, such as that diagrammatically shown in Fig. 1. Data is entered into system 10 via any suitable type of user interface 16, and may be stored in memory 12, which may be
any suitable type of computer readable and programmable memory and is preferably a non- transitory, computer readable storage medium. Calculations are performed by processor 14, which may be any suitable type of computer processor and may be displayed to the user on display 18, which may be any suitable type of computer display. Sensor data is collected from solar irradiance sensors 20, such as, for example, satellite and sky cameras, radiometric sensors, pyrheliometers, pyranometers and the like.
Processor 14 may be associated with, or incorporated into, any suitable type of computing device, for example, a personal computer or a programmable logic controller. The display 18, the processor 14, the memory 12 and any associated computer readable recording media are in communication with one another by any suitable type of data bus, as is well known in the art.
Examples of computer-readable recording media include non-transitory storage media, a magnetic recording apparatus, an optical disk, a magneto-optical disk, and/or a semiconductor memory (for example, RAM, ROM, etc.). Examples of magnetic recording apparatus that may be used in addition to memory 112, or in place of memory 12, include a hard disk device (HDD), a flexible disk (FD), and a magnetic tape (MT). Examples of the optical disk include a DVD (Digital Versatile Disc), a DVD-RAM, a CD-ROM (Compact Disc-Read Only Memory), and a CD-R (Recordable)/RW. It should be understood that non- transitory computer-readable storage media include all computer-readable media, with the sole exception being a transitory, propagating signal.
In the above method, the Persistence technique and the Support Vector Regression (SVR) technique may be used as the forecasting methods. As an alternative, the Persistence model may be replaced by autoregressive (AR) models, particularly for the prediction of solar radiation in the short term (i.e., "now-casting") using ground measurements, such as radiometric measurements. This allows for temporal steps of predictions of 1, 5 and 10 minutes. The temporal horizon in all three cases is 15 temporal steps ahead. In the below, due to non-stationary behavior, global solar irradiance has been transformed to a normalized clearness index Kj. The clearness index is defined as the ratio between ground measured global solar irradiance and extraterrestrial solar irradiance, including a correction for air mass.
Autoregressive models generate lineal predictions from the input and are also known as Infinite Impulse Response Filters (IIRFs). The notation AR(/?) refers to the autoregressive model of order p. The AR(p) is expressed by the following equation:
Xt — C +∑ =1 CliXt-i + Et, where at are the parameters of the model, c is a constant which represents the mean value of the time series, and st is a white noise signal. As will be discussed in greater detail below, twenty different autoregressive models AR(/?) have been tested, with order p = 1,..., 20 for 1, 5 and 10 minutes average time series. Exemplary data are from the Qatar Environment and Energy Research Institute (QEERI), which has been operating a high precision solar radiation monitoring station since the end of November 2012 in Education City, Doha (25.33°N, 51.43Έ). The station is equipped with a solar tracker with a sun sensor kit, for improved tracking accuracy, and a shading ball assembly for diffuse measurements. Mounted on the sun tracker are one first class pyrheliometer for measuring Direct Normal Irradiance (DNI), and two secondary standard pyranometers (one of them shaded) for Global Horizontal Irradiance (GHI) and Diffuse Horizontal Irradiance (DHI) measurements. Both pyranometers are fitted with ventilation units. Data from the monitoring station are sampled every second and recorded as minute averages in W/m2.
Figs. 2 A and 2B show autocorrelation and partial autocorrelation of a one-minute time series used in the alternative embodiment of the method of forecasting for solar-based power systems, respectively. Similarly, Figs. 3A and 3B show autocorrelation and partial autocorrelation of a five-minute time series, respectively, and Figs. 4A and 4B show autocorrelation and partial autocorrelation of a ten- minute time series. The metrics to measure the errors rates of the models are the mean bias deviation (MBD), the relative root mean squared deviation (RMSD) and its relative values rMBD and rRMSD normalized with the mean value of the observed variable to predict for the period under validation.
The present models are compared with a baseline model, with the intention of measuring the improvement achieved. The basic model chosen is the persistence model (PER) because it is the most extended model to contrast new proposed models. As described above, persistence is based on the assumption that the value for the next temporal step is the same as the present value: xt+k = xt, where xt+k is the prediction for the next k steps, and xt is the observation at the temporal instant t.
Fig. 5 is a graph comparing relative root mean squared deviation (RMSD) of the autocorrelation (AR) model against the persistence (PER) model, and a combination autocorrelation model in the alternative embodiment of the method of forecasting for solar- based power systems, specifically for each horizon of prediction applied to one-minute average time series data (taken from January of 2014). The new "AR combined" model
(AR_COMB_MIN) is detailed below. Figs. 6 and 7 show similar comparisons for data sets recorded in April of 2014 and June of 2014, respectively.
Analysis of this data shows that there is a dependence on the horizon of prediction and the complexity of the AR models which provide the minimum error. For each horizon n, the best results in terms of minimum rRMSD are obtained with an AR model of order
(n + l) [AR(n + 1)] . This is the "AR combined" model (AR_COMB_MIN) noted above. The same behavior is also observed in the 5 and 10 minutes time series. Table 6 shows MBD, RMSD, rMBD and rRMSD results for the AR, PER and AR-combined models with the one-minute time series from 2014.
Table 6: MBD, RMSD, rMBD and rRMSD Results for AR, PER and AR-combined
Models with One-minute Time Series
Model MBD RMSD rMBD(%) rRMSD(%)
AR(1) -0.01 0.06 -0.84 9.44
AR(2) -0.01 0.06 -0.81 8.76
AR(3) 0.00 0.06 -0.59 8.26
AR(4) 0.00 0.05 -0.50 7.91
AR(5) 0.00 0.05 -0.43 7.63
AR(6) 0.00 0.05 -0.39 7.40
AR(7) 0.00 0.05 -0.35 7.24
AR(8) 0.00 0.05 -0.31 7.12
AR(9) 0.00 0.05 -0.28 7.01
AR(10) 0.00 0.05 -0.26 6.94
AR(l l) 0.00 0.05 -0.24 6.90
AR(12) 0.00 0.05 -0.23 6.91
AR(13) 0.00 0.05 -0.21 6.99
AR(14) 0.00 0.05 -0.21 7.07
AR(15) 0.00 0.05 -0.20 7.24
AR(16) 0.00 0.05 -0.18 7.51
AR(17) 0.00 0.05 -0.18 7.78
AR(18) 0.00 0.05 -0.17 8.04
AR(19) 0.00 0.06 -0.17 8.09
AR(20) 0.00 0.05 -0.17 8.04
PER 0.00 0.06 0.13 9.42
AR_COMB_MIN 0.00 0.02 -0.26 3.30
Fig. 8 shows model rRMSD results for the one-minute time series for each temporal horizon of prediction. The above data shows that the AR-combined model is better than any other model, including the PER baseline model. The difference is quite considerable during the entire year, except in June. The lower limit of monthly rRMSD is around 6%. Fig. 9 is a graph comparing relative root mean squared deviation (RMSD) of the autocorrelation (AR)
model against the persistence (PER) model, and a combination autocorrelation model in the alternative embodiment of the method of forecasting for solar-based power systems, specifically for each horizon of prediction applied to the five-minute average time series data (taken from January of 2014). Figs. 10 and 11 show similar comparisons for data sets recorded in June of 2014 and November of 2014, respectively.
Table 7 shows MBD, RMSD, rMBD and rRMSD results for the AR, PER and AR- combined models with the five-minute time series from 2014.
MBD, RMSD, rMBD and rRMSD Results for AR, PER and AR-combined
Models with Five-minute Time Series
Model MBD RMSD rMBD(%) rRMSD(%)
AR(1) -0.01 0.09 -0.96 13.38
AR(2) 0.00 0.08 -0.56 12.46
AR(3) 0.00 0.08 -0.14 11.89
AR(4) 0.00 0.08 0.05 11.53
AR(5) 0.00 0.08 0.19 11.28
AR(6) 0.00 0.08 0.26 11.14
AR(7) 0.00 0.07 0.32 11.08
AR(8) 0.00 0.08 0.34 11.11
AR(9) 0.00 0.08 0.36 11.22
AR(10) 0.00 0.08 0.37 11.38
AR(ll) 0.00 0.08 0.38 11.55
AR(12) 0.00 0.08 0.38 11.77
AR(13) 0.00 0.08 0.38 12.01
AR(14) 0.00 0.08 0.38 12.29
AR(15) 0.00 0.09 0.39 12.62
AR(16) 0.00 0.09 0.41 13.02
AR(17) 0.00 0.09 0.40 13.41
AR(18) 0.00 0.09 0.39 13.82
AR(19) 0.00 0.10 0.38 14.21
AR(20) 0.00 0.10 0.39 14.58
PER 0.01 0.09 1.30 13.40
AR_COMB_MIN 0.00 0.05 0.33 7.72
Fig. 12 shows model rRMSD results for the five-minute time series for each temporal horizon of prediction. From the above data, one can conclude that the AR-combined model outperforms the other models and the PER baseline model for the case of a five-minute average time series. The difference is quite considerable during the entire year, except in June. The lower limit of monthly rRMSD is around 6%.
Fig. 13 is a graph comparing relative root mean squared deviation (RMSD) of the autocorrelation (AR) model against the persistence (PER) model, and the combination
autocorrelation model in the alternative embodiment of the method of forecasting for solar- based power systems, specifically for each horizon of prediction applied to the ten- minute average time series data (taken from February of 2014). Figs. 14 and 15 show similar comparisons for data sets recorded in August of 2014 and December of 2014, respectively.
Table 8 shows MBD, RMSD, rMBD and rRMSD results for the AR, PER and AR- combined models with the ten- minute time series from 2014.
MBD, RMSD, rMBD and rRMSD Results for AR, PER and AR-combined
Models with Ten-minute Time Series
Model MBD RMSD rMBD(%) rRMSD(%)
AR(1) 0.00 0.09 -0.23 16.26
AR(2) 0.00 0.08 0.26 15.19
AR(3) 0.01 0.08 0.72 14.60
AR(4) 0.01 0.08 0.89 14.23
AR(5) 0.01 0.08 0.95 14.02
AR(6) 0.01 0.08 0.96 13.93
AR(7) 0.01 0.08 0.90 13.90
AR(8) 0.01 0.08 0.88 13.95
AR(9) 0.01 0.08 0.85 14.04
AR(10) 0.01 0.08 0.77 14.18
AR(l l) 0.01 0.08 0.73 14.41
AR(12) 0.01 0.08 0.68 14.70
AR(13) 0.01 0.08 0.65 15.07
AR(14) 0.01 0.09 0.61 15.48
AR(15) 0.00 0.09 0.60 15.93
AR(16) 0.00 0.09 0.61 16.40
AR(17) 0.00 0.09 0.62 16.89
AR(18) 0.00 0.10 0.60 17.27
AR(19) 0.00 0.10 0.58 17.69
AR(20) 0.00 0.10 0.55 18.10
PER 0.02 0.10 2.62 16.39
AR_COMB_MIN 0.01 0.06 0.61 10.18
Fig. 16 shows model rRMSD results for the ten-minute time series for each temporal horizon of prediction. From the above data, one can conclude that the AR-combined model is better than any AR model and the PER model, also for the case of 10 minute average time series. The difference is quite considerable during the entire year, except in June. This is due to the fact that June is a month where there are almost no clouds in Qatar and almost no impact in lower variability of solar radiation. The lower limit of monthly rRMSD is around 9%.
It is to be understood that the present invention is not limited to the embodiments described above, but encompasses any and all embodiments within the scope of the following claims.
Claims
1. A method of forecasting for solar-based power systems, comprising the steps of:
(a) measuring solar irradiance parameters with sensors for a defined geographical region over predetermined time intervals to form a data set;
(b) selecting a window size defining a number of past measurements and future forecast predictions to be made from the number of past measurements;
(c) partitioning the data set into successive and adjacent time series training data sequences of the selected window size;
(d) applying a plurality of forecasting methods to the time series training data sequences to obtain future forecast predictions from each of the forecasting methods;
(e) comparing the future forecast predictions of each of the forecasting methods to measured data to obtain a corresponding error rate associated with each of the methods, given the time series training data sequences;
(f) assigning the forecasting method with the lowest error rate as the forecasting class for the time series training data sequences;
(g) repeating steps (a) through (f) to train a classifier to determine an optimal forecasting class for different time series training data sequences;
(h) using the sensors to measure current solar irradiance parameters;
(i) using the classifier to determine the optimal forecasting class for the current solar irradiance parameters;
(j) making future forecast predictions from the current solar irradiance parameters using the optimal forecasting class;
(k) predicting solar-based power system demands and generating capacities based upon the future forecast predictions made in step (j); and
(1) making adjustments in the solar-based power system demands and stored energy in order to maintain a substantially constant voltage supply for the geographic region.
2. The method of forecasting for solar-based power systems as recited in claim 1, wherein the step of measuring solar irradiance parameters comprises measuring direct normal irradiance (DNI), global horizontal irradiance (GHI) and diffuse horizontal irradiance (DHI).
3. The method of forecasting for solar-based power systems as recited in claim 2, wherein the step of measuring solar irradiance parameters comprises measuring the solar irradiance parameters in one minute intervals.
4. The method of forecasting for solar-based power systems as recited in claim 1, further comprising the step of applying data filtering to the data set generated in step (a).
5. The method of forecasting for solar-based power systems as recited in claim 1, wherein the plurality of forecasting methods comprise a persistence method and a support vector regression method.
6. A method of forecasting for solar-based power systems, comprising the steps of: measuring solar irradiance parameters with sensors for a defined geographical region over predetermined time intervals to form a data set;
selecting a window size defining a number of past measurements and future forecast predictions to be made from the number of past measurements;
partitioning the data set into successive and adjacent time series training data sequences of the selected window size;
applying an autoregressive forecasting method to the time series training data sequences to obtain future forecast predictions;
predicting solar-based power system demands and generating capacities based upon the future forecast predictions; and
making adjustments in the solar-based power system demands and stored energy in order to maintain a substantially constant voltage supply for the geographic region.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP16815281.7A EP3314751B1 (en) | 2015-06-23 | 2016-06-23 | Method of forecasting for solar-based power systems |
US15/738,102 US11063555B2 (en) | 2015-06-23 | 2016-06-23 | Method of forecasting for solar-based power systems |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US201562183705P | 2015-06-23 | 2015-06-23 | |
US62/183,705 | 2015-06-23 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2016210102A1 true WO2016210102A1 (en) | 2016-12-29 |
Family
ID=57585675
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US2016/038977 WO2016210102A1 (en) | 2015-06-23 | 2016-06-23 | Method of forecasting for solar-based power systems |
Country Status (3)
Country | Link |
---|---|
US (1) | US11063555B2 (en) |
EP (1) | EP3314751B1 (en) |
WO (1) | WO2016210102A1 (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107609774A (en) * | 2017-09-11 | 2018-01-19 | 华北电力大学 | A kind of photovoltaic power Forecasting Methodology based on mind evolutionary Optimization of Wavelet neutral net |
CN109063938A (en) * | 2018-10-30 | 2018-12-21 | 浙江工商大学 | Air Quality Forecast method based on PSODE-BP neural network |
WO2019021438A1 (en) * | 2017-07-27 | 2019-01-31 | 三菱電機株式会社 | Solar power generation amount prediction device, solar power generation amount prediction system, prediction method, and program |
WO2019130718A1 (en) * | 2017-12-28 | 2019-07-04 | 住友電気工業株式会社 | Determination device, photovoltaic power generation system, determination method, and determination program |
CN110361090A (en) * | 2019-06-20 | 2019-10-22 | 广东工业大学 | The following illuminance prediction technique based on photovoltaic array sensors association |
US10732319B2 (en) | 2017-08-30 | 2020-08-04 | International Business Machines Corporation | Forecasting solar power output |
CN112200377A (en) * | 2020-10-16 | 2021-01-08 | 国能日新科技股份有限公司 | Photovoltaic medium-long term power generation capacity forecasting method and device based on SARIMAX model |
WO2021096429A1 (en) * | 2019-11-14 | 2021-05-20 | Envision Digital International Pte. Ltd. | Method for processing irradiation forecast, method for training stacked generalization model, and apparatuses thereof |
CN116451598A (en) * | 2023-06-20 | 2023-07-18 | 杭州经纬信息技术股份有限公司 | Solar irradiance prediction method based on denoising diffusion probability model |
Families Citing this family (24)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2013181408A2 (en) * | 2012-05-30 | 2013-12-05 | Neo Virtus Engineering, Inc. | Method and apparatus for forecasting solar radiation and solar power production using synthetic irradiance imaging |
US20180150857A1 (en) * | 2016-11-30 | 2018-05-31 | Electronics And Telecommunications Research Institute | Apparatus and method for processing tender of aggregate power plant |
US11615476B2 (en) * | 2016-12-22 | 2023-03-28 | Sony Corporation | Information processing device and method |
US10922634B2 (en) * | 2017-05-26 | 2021-02-16 | General Electric Company | Determining compliance of a target asset to at least one defined parameter based on a simulated transient response capability of the target asset and as a function of physical operation data measured during an actual defined event |
CN107341569B (en) * | 2017-06-26 | 2020-04-24 | 清华大学 | Photovoltaic power prediction method combining photovoltaic power physical model and data driving |
US10579123B2 (en) * | 2018-01-12 | 2020-03-03 | Samsara Networks Inc. | Adaptive power management in a battery powered system based on expected solar energy levels |
JP6586184B2 (en) * | 2018-03-13 | 2019-10-02 | 株式会社日立製作所 | Data analysis support device and data analysis support method |
US10956132B1 (en) * | 2018-06-11 | 2021-03-23 | Amazon Technologies, Inc. | Unified code and data management for model development |
CN109472283B (en) * | 2018-09-13 | 2022-02-01 | 中国科学院计算机网络信息中心 | Dangerous weather prediction method and device based on multiple incremental regression tree model |
US11347629B2 (en) * | 2018-10-31 | 2022-05-31 | Dell Products L.P. | Forecasting a quality of a software release using machine learning |
US11693378B2 (en) * | 2019-03-01 | 2023-07-04 | Alliance For Sustainable Energy, Llc | Image-based solar estimates |
FR3095067A1 (en) * | 2019-04-11 | 2020-10-16 | Total Solar | Photovoltaic energy production evaluation method and evaluation and management unit implementing the method |
CN110086426B (en) * | 2019-04-28 | 2020-06-19 | 西北核技术研究所 | Method for rapidly acquiring I-V curve of two-end type laminated solar cell |
CN110378507B (en) * | 2019-05-22 | 2023-05-12 | 国网辽宁省电力有限公司大连供电公司 | High-temperature high-capacity heat storage system management and control method based on weather prediction |
CN110781458B (en) * | 2019-10-30 | 2023-04-07 | 云南师范大学 | Method for predicting surface solar irradiance based on mixed regression model |
US11823083B2 (en) * | 2019-11-08 | 2023-11-21 | International Business Machines Corporation | N-steps-ahead prediction based on discounted sum of m-th order differences |
CN111242345A (en) * | 2019-12-26 | 2020-06-05 | 浙江大学 | Nuclear power unit electric power prediction method based on cluster analysis and random forest regression |
CN111210093B (en) * | 2020-03-05 | 2023-05-09 | 重庆森鑫炬科技有限公司 | Daily water consumption prediction method based on big data |
KR102346188B1 (en) * | 2020-04-08 | 2021-12-31 | 상명대학교산학협력단 | method for forecasting power demanding and apparatus adopting the method |
CN111814398B (en) * | 2020-07-08 | 2023-09-29 | 国网河北省电力有限公司 | Map-based earth surface solar radiance prediction method integrating space-time attention |
CN113218092A (en) * | 2021-05-11 | 2021-08-06 | 沈阳建筑大学 | Solar heat collector coupling system operation method based on temperature prediction |
CN113361946B (en) * | 2021-06-23 | 2023-09-22 | 云南电网有限责任公司电力科学研究院 | Power quality assessment method and device based on distributed photovoltaic grid-connected system |
CN115629431B (en) * | 2022-12-22 | 2023-03-14 | 成都数之联科技股份有限公司 | Water vapor content prediction method, device, equipment and medium |
CN116154768B (en) * | 2023-04-14 | 2023-06-27 | 南方电网数字电网研究院有限公司 | Power interval prediction method adopting point prediction error empirical distribution inverse transformation |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2007047048A2 (en) * | 2005-10-18 | 2007-04-26 | Gm Global Technology Operations, Inc. | Improved solar photovoltaic output for cloudy conditions with a solar tracking system |
US20120191351A1 (en) * | 2007-02-12 | 2012-07-26 | Locus Energy, Llc | Estimating solar irradiance components from plane of array irradiance and global horizontal irradiance |
US20130211722A1 (en) * | 2011-07-25 | 2013-08-15 | Clean Power Research, L.L.C. | Computer-Implemented System And Method For Bounding Accuracy On A Forecast Of Photovoltaic Fleet Power Generation |
US20140136178A1 (en) * | 2012-11-15 | 2014-05-15 | Power Analytics Corporation | Systems and methods for model-based solar power management |
JP2015005641A (en) * | 2013-06-21 | 2015-01-08 | 株式会社東芝 | Predicting system, predicting apparatus and predicting method |
Family Cites Families (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE602004010281T2 (en) * | 2003-08-20 | 2008-10-02 | New Energy Options, Inc., Littleton | METHOD AND SYSTEM FOR PREDICTING SOLAR ENERGY PRODUCTION |
US20100100328A1 (en) | 2008-06-17 | 2010-04-22 | Moore John S | System and Method for Generating a Cloud Type and Coverage Prediction Database |
US8295989B2 (en) | 2009-02-03 | 2012-10-23 | ETM Electromatic, Inc. | Local power tracking for dynamic power management in weather-sensitive power systems |
WO2011158363A1 (en) * | 2010-06-17 | 2011-12-22 | 株式会社四国総合研究所 | Solar radiation intensity prediction system and photovoltaic power generation output prediction system |
US8682585B1 (en) * | 2011-07-25 | 2014-03-25 | Clean Power Research, L.L.C. | Computer-implemented system and method for inferring operational specifications of a photovoltaic power generation system |
US8645304B2 (en) | 2011-08-19 | 2014-02-04 | International Business Machines Corporation | Change point detection in causal modeling |
ES2644528T3 (en) | 2011-12-21 | 2017-11-29 | Siemens Aktiengesellschaft | Method for the computer-aided determination of the use of electric power produced by a power generation plant, particularly a renewable power generation plant |
US9633359B2 (en) | 2012-08-10 | 2017-04-25 | Itron, Inc. | Near-term data filtering, smoothing and load forecasting |
WO2014075108A2 (en) | 2012-11-09 | 2014-05-15 | The Trustees Of Columbia University In The City Of New York | Forecasting system using machine learning and ensemble methods |
US20140278107A1 (en) | 2013-03-12 | 2014-09-18 | Locus Energy, Llc | Methods and systems for real-time solar forecasting incorporating a ground network |
US20140278108A1 (en) * | 2013-03-13 | 2014-09-18 | Locus Energy, Llc | Methods and Systems for Optical Flow Modeling Applications for Wind and Solar Irradiance Forecasting |
US10168448B2 (en) | 2013-04-30 | 2019-01-01 | International Business Machines Corporation | Machine learning approach for analysis and prediction of cloud particle size and shape distribution |
US9134458B2 (en) * | 2013-06-25 | 2015-09-15 | General Electric Company | Prediction of solar obscuration events based on detection of spectral distribution shifts caused by approaching clouds |
AU2014389497B2 (en) * | 2014-04-04 | 2017-09-07 | Siemens Aktiengesellschaft | Combing multiple trending models for photovoltaic plant output forecasting |
-
2016
- 2016-06-23 EP EP16815281.7A patent/EP3314751B1/en active Active
- 2016-06-23 US US15/738,102 patent/US11063555B2/en active Active
- 2016-06-23 WO PCT/US2016/038977 patent/WO2016210102A1/en active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2007047048A2 (en) * | 2005-10-18 | 2007-04-26 | Gm Global Technology Operations, Inc. | Improved solar photovoltaic output for cloudy conditions with a solar tracking system |
US20120191351A1 (en) * | 2007-02-12 | 2012-07-26 | Locus Energy, Llc | Estimating solar irradiance components from plane of array irradiance and global horizontal irradiance |
US20130211722A1 (en) * | 2011-07-25 | 2013-08-15 | Clean Power Research, L.L.C. | Computer-Implemented System And Method For Bounding Accuracy On A Forecast Of Photovoltaic Fleet Power Generation |
US20140136178A1 (en) * | 2012-11-15 | 2014-05-15 | Power Analytics Corporation | Systems and methods for model-based solar power management |
JP2015005641A (en) * | 2013-06-21 | 2015-01-08 | 株式会社東芝 | Predicting system, predicting apparatus and predicting method |
Non-Patent Citations (1)
Title |
---|
See also references of EP3314751A4 * |
Cited By (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019021438A1 (en) * | 2017-07-27 | 2019-01-31 | 三菱電機株式会社 | Solar power generation amount prediction device, solar power generation amount prediction system, prediction method, and program |
JPWO2019021438A1 (en) * | 2017-07-27 | 2020-04-02 | 三菱電機株式会社 | Photovoltaic power generation prediction device, photovoltaic power generation prediction system, prediction method and program |
US10732319B2 (en) | 2017-08-30 | 2020-08-04 | International Business Machines Corporation | Forecasting solar power output |
CN107609774A (en) * | 2017-09-11 | 2018-01-19 | 华北电力大学 | A kind of photovoltaic power Forecasting Methodology based on mind evolutionary Optimization of Wavelet neutral net |
CN107609774B (en) * | 2017-09-11 | 2023-12-08 | 华北电力大学 | Photovoltaic power prediction method for optimizing wavelet neural network based on thought evolution algorithm |
JP7188399B2 (en) | 2017-12-28 | 2022-12-13 | 住友電気工業株式会社 | Determination device, photovoltaic power generation system, determination method and determination program |
WO2019130718A1 (en) * | 2017-12-28 | 2019-07-04 | 住友電気工業株式会社 | Determination device, photovoltaic power generation system, determination method, and determination program |
JPWO2019130718A1 (en) * | 2017-12-28 | 2020-12-24 | 住友電気工業株式会社 | Judgment device, photovoltaic power generation system, judgment method and judgment program |
CN109063938A (en) * | 2018-10-30 | 2018-12-21 | 浙江工商大学 | Air Quality Forecast method based on PSODE-BP neural network |
CN110361090A (en) * | 2019-06-20 | 2019-10-22 | 广东工业大学 | The following illuminance prediction technique based on photovoltaic array sensors association |
CN110361090B (en) * | 2019-06-20 | 2021-05-25 | 广东工业大学 | Future illuminance prediction method based on relevance of photovoltaic array sensor |
KR20220088946A (en) * | 2019-11-14 | 2022-06-28 | 엔비전 디지털 인터내셔널 피티이 리미티드 | Radial prediction processing method, stack generalization model training method and apparatus |
WO2021096429A1 (en) * | 2019-11-14 | 2021-05-20 | Envision Digital International Pte. Ltd. | Method for processing irradiation forecast, method for training stacked generalization model, and apparatuses thereof |
KR102500939B1 (en) | 2019-11-14 | 2023-02-17 | 엔비전 디지털 인터내셔널 피티이 리미티드 | Radiation prediction processing method, stack generalization model training method and apparatus |
US11842303B2 (en) | 2019-11-14 | 2023-12-12 | Envision Digital International Pte. Ltd. | Method for processing irradiation forecast, method for training stacked generalization model, and apparatuses thereof |
CN112200377A (en) * | 2020-10-16 | 2021-01-08 | 国能日新科技股份有限公司 | Photovoltaic medium-long term power generation capacity forecasting method and device based on SARIMAX model |
CN116451598A (en) * | 2023-06-20 | 2023-07-18 | 杭州经纬信息技术股份有限公司 | Solar irradiance prediction method based on denoising diffusion probability model |
CN116451598B (en) * | 2023-06-20 | 2023-09-05 | 杭州经纬信息技术股份有限公司 | Solar Irradiance Prediction Method Based on Denoising Diffusion Probability Model |
Also Published As
Publication number | Publication date |
---|---|
EP3314751A4 (en) | 2019-03-20 |
US20180175790A1 (en) | 2018-06-21 |
EP3314751A1 (en) | 2018-05-02 |
EP3314751B1 (en) | 2020-09-02 |
US11063555B2 (en) | 2021-07-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11063555B2 (en) | Method of forecasting for solar-based power systems | |
Sharma et al. | Forecasting daily global solar irradiance generation using machine learning | |
Carneiro et al. | Review on photovoltaic power and solar resource forecasting: current status and trends | |
Abuella et al. | Solar power forecasting using support vector regression | |
US11022720B2 (en) | System for forecasting renewable energy generation | |
EP3576029A1 (en) | Method and device for determining energy system operating scenarios | |
Parvez et al. | Multi-layer perceptron based photovoltaic forecasting for rooftop pv applications in smart grid | |
Mathe et al. | PVNet: A LRCN architecture for spatio-temporal photovoltaic PowerForecasting from numerical weather prediction | |
Badosa et al. | Day-ahead probabilistic forecast of solar irradiance: a Stochastic Differential Equation approach | |
Alonso-Montesinos et al. | Impact of DNI forecasting on CSP tower plant power production | |
Hocaoğlu | Novel analytical hourly solar radiation models for photovoltaic based system sizing algorithms | |
Isaksson et al. | Solar power forecasting with machine learning techniques | |
Larrañeta et al. | Methodology to synthetically downscale DNI time series from 1-h to 1-min temporal resolution with geographic flexibility | |
Terhag et al. | Optimization of cleaning strategies based on ANN algorithms assessing the benefit of soiling rate forecasts | |
Assouline et al. | Estimation of large-scale solar rooftop PV potential for smart grid integration: A methodological review | |
US20220352714A1 (en) | System for estimating renewable energy generation quantity in real-time | |
Haupt | Short-range forecasting for energy | |
Paulescu et al. | Intra-hour PV power forecasting based on sky imagery | |
Suksamosorn et al. | Post-processing of NWP forecasts using Kalman filtering with operational constraints for day-ahead solar power forecasting in Thailand | |
Anvari-Moghaddam et al. | Feasibility study of a novel methodology for solar radiation prediction on an hourly time scale: A case study in Plymouth, United Kingdom | |
Al-Hilfi et al. | Enhancing the estimation of the overall produced power by several adjacent photovoltaic systems using existing correlational factors | |
Sanfilippo | Solar nowcasting | |
KR20230000751A (en) | Method of prediction for solar power generation and computing device for performing the method | |
Anand et al. | Analysis of the quality of long-term synthetic solar radiation data generated from stochastic models | |
Nunnari | Forecasting the Class of Daily Clearness Index for PV Applications. |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 16815281 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 15738102 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |