WO2014178877A1 - Method for using posterior information to improve forecast capabilities of sequential kernel regression models - Google Patents
Method for using posterior information to improve forecast capabilities of sequential kernel regression models Download PDFInfo
- Publication number
- WO2014178877A1 WO2014178877A1 PCT/US2013/039456 US2013039456W WO2014178877A1 WO 2014178877 A1 WO2014178877 A1 WO 2014178877A1 US 2013039456 W US2013039456 W US 2013039456W WO 2014178877 A1 WO2014178877 A1 WO 2014178877A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- vector
- future
- vector sequences
- data
- matching
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
- G05B23/0224—Process history based detection method, e.g. whereby history implies the availability of large amounts of data
- G05B23/024—Quantitative history assessment, e.g. mathematical relationships between available data; Functions therefor; Principal component analysis [PCA]; Partial least square [PLS]; Statistical classifiers, e.g. Bayesian networks, linear regression or correlation analysis; Neural networks
Definitions
- the present invention generally relates to kernel regression modeling, and more specifically to use of the kernel regression modeling to forecast future behavior of a monitored system.
- Sequential Kernel Regression is an empirical modeling method that merges the powerful multivariate pattern recognition capabilities of kernel regression with explicit pattern matching through time to provide unique capabilities for forecasting multiple sensor trends across wide time scales.
- SKR utilizes a pattern matching technique that matches an input pattern sequence, composed of a sequence of recent observation vectors ending with the current observation, to arrays of contiguous observation vectors in the training matrix to identify periods of past operation most similar to current system behavior.
- Each of the identified periods of past operation has its own future, represented by observation vectors that succeed the matched vectors.
- the modeling technique produces forecasts of future system behavior. In this manner, the modeling technique utilizes prior information, as represented by the input pattern sequence, to produce forecasts of future behavior.
- Embodiments of the present invention overcome the aforementioned deficiencies noted in the conventional modeling methods.
- An aspect of embodiments of the present invention is to provide a method for improving future forecast by using known posterior information for a set of candidate conditions.
- a method for forecasting future behaviors of a system includes receiving, by a forecasting server, a plurality of data, forming a sequence of observed vectors with the plurality of data, comparing the observed sequence of vectors with a plurality of stored vector sequences.
- the method also includes generating a plurality of matching vectors according to a comparison result, comparing a plurality of future conditions with posterior information associated with each matching vector.
- a sequence of forecast vectors is obtained with future behaviors based on comparison of the plurality of future conditions with the posterior information from each matching vector.
- an apparatus for improving forecast of future behavior of a system.
- the apparatus includes a monitoring port interface for receiving monitoring data from a monitored system and a vector sequence comparator for forming a sequence of observed vectors with the monitoring data.
- the observed vector sequence is compared with a plurality of stored vector sequences.
- the apparatus also includes generating a plurality of matching vectors based on a comparison result, comparing the plurality of matching vectors with a future condition, and generating a sequence of forecast vectors based on comparison between the plurality of matching vectors with the future condition.
- a non-transient computer-readable storage medium having computer executable instructions, when executed by a computer, causes the computer to perform certain operations.
- the operations include receiving a plurality of data, forming a sequence of observed vectors with the plurality of data, comparing the observed vector sequence with a plurality of stored vector sequences.
- the operation also includes generating a plurality of matching vectors according to a comparison result, and comparing a plurality of future conditions with posterior information associated with each matching vector.
- a sequence of forecast vectors is obtained with future behaviors based on comparison of the plurality of future conditions with the posterior information from each matching vector.
- FIG. 1 shows architecture of a forecast system according to the present invention
- FIG. 2 illustrates a first step in the forecast method according to the invention
- FIG. 3 illustrates a second step 300 in the forecast method using posterior information according to the present invention
- FIG. 4 is a flowchart of the forecast method according to the present invention.
- FIG. 5 is a chart illustrating comparative forecasting results
- FIG. 6 illustrates architecture of a forecasting device according to the present invention.
- FIG. 1 is architecture 100 of a forecasting and prognostic system according to the invention.
- a device 102 is a device to be monitored and is set with many sensors for collecting the data at each sensing point.
- the collected data are sent to the forecasting server 104.
- the forecasting server 104 compares the collected data with data from a database 106 and produces a forecast that is displayed on a user device 108 to the user 110.
- the forecasting server 104 compares and produces the forecast by using Sequential Kernel Regression Modeling technique.
- enhanced SKR modeling significantly improves forecasting of modeled parameters over the conventional SKR modeling technologies.
- Both enhanced and conventional SKR merge the powerful cross-parameter pattern recognition capabilities of kernel regression with explicit pattern matching through time to provide unique capabilities for extrapolating sensor trends across wide time scales.
- Conventional SKR utilizes a pattern matching technique that matches the sequence of recent observation vectors ending with the current observation to equal-sized sequences from past operation of the modeled system.
- the most recent sequence of vectors called the input pattern sequence, is an array containing data from the set of model variables collected during a contiguous set of time points ending with the most recent time point.
- a set of observation vectors that precede the current observation vector is referred to herein as the look back window. The number of vectors in it is called the look back count.
- the conventional SKR compares the input pattern sequence to arrays of contiguous observation vectors in the training matrix to identify periods of past operation most similar to current system behavior.
- Each of the periods in the past identified by the pattern matching process has their own future, represented by observation vectors from time points that succeed the matched vectors.
- the conventional SKR technique collects a set of succeeding vectors, referred to herein as the look ahead window, from each of the time periods it has matched to the input pattern sequence.
- the SKR modeling mathematical expression operates on the data in the input pattern sequence and the data in the matched periods from the past to calculate an array of estimate vectors.
- enhanced SKR utilizes pattern matching to identify periods in the past that are similar to data in the input pattern sequence. But where enhanced SKR departs from conventional SKE, it is that the pattern matching algorithm is augmented by expectations of future system response to identify periods in the past that are similar to the input pattern sequence and the expected future behavior.
- Each estimate vector contains an estimate of the expected behavior of all modeled variables at a point in time.
- the vectors in the estimate array correspond to the time points in the look back window, the current time point, and the time points in the look ahead window.
- the data in the look ahead window are estimates of the expected future behavior of the modeled variables.
- enhanced SKR incorporates posterior information to improve upon the pattern matching process of conventional SKR.
- posterior information is data about future events that has been produced by means independent of those employed by the sequential kernel regression model.
- An SKR model of a wind turbine includes data from sensors that measure site wind speed and measurements of the power produced by individual and/or groups of turbines.
- Another example of posterior information is econometric forecasts of market conditions for use with SKR models of financial time series data, such as stock indexes, equity prices, and interest rates.
- the embodiments provide a new means to perform scenario analysis - the process of analyzing the effect of possible future events on the behavior of financial instruments.
- Yet another example of posterior information is the end state of an equipment fault for use with SKR models of industrial equipment.
- the embodiments provide a new means of performing prognostics by estimating the probability of various fault scenarios and remaining useful life of the equipment given current evidence of equipment degradation.
- an SKR model of a group of wind turbines co-located at a single geographic site was modified to use meteorological wind speed forecast data.
- Sensor data from a site containing 52 wind turbines were analyzed with an SKR model.
- Sensor data included in the model consisted of the power produced by 8 turbine groups, the total site power, and a single site wind speed signal. These data were augmented with a set of data files that contained 48 hour forecasts of various meteorological parameters, including wind speed and direction for the site.
- a meteorological forecast file was provided for each hour of wind site operation over a 17 month period.
- the wind speed data were extracted from the forecast files, providing a rolling 48 hour forecast of wind conditions. Wind speed forecast data were not directly included in the model. Instead, the embodiments operate on the forecast wind speed data to improve pattern matching between current and past periods of operation. As an aspect of the embodiments, two tests were used to improve pattern matching.
- pattern matching of the forecast wind speed to wind speed data in the training matrix is the only method used to identify periods of past operation most similar to current system behavior.
- FIG. 5 is an illustration of standard error for the forecast of a total power signal as a function of a look ahead count. Results from three calculations are illustrated in FIG. 5. Results from calculations using the two tests, aspects of the embodiments, are depicted by lines 502 and 504. Results obtained using the conventional SKR algorithm are depicted by line 506. Study of FIG. 5 reveals that by using meteorological forecast data to improve pattern matching, in accordance with the embodiments, the long-term ability of the SKR models to forecast site behavior is substantially improved.
- the conventional SKR pattern matching between an input pattern sequence and sequences of observation vectors in a training matrix is augmented. This augmentation occurs by pattern matching of the forecast wind speed to wind speed data from observation vectors in the training matrix that succeed vectors matched by the conventional SKR.
- the enhanced SKR improves pattern matching, ultimately providing more accurate forecasts of site -wide turbine power production.
- FIG. 2 is an illustration 200 of a first step in the enhanced SKR modeling according to embodiments of the present invention.
- Data from recent observations of the monitored system 102 is put into a vector sequence format.
- the observed vector sequence 202 of the recent observations is compared with a plurality of past vector sequences (a.k.a. stored vector sequences) 204 retrieved from the database 106.
- the vector sequences retrieved from the database 106 are data collected from past observations of the monitored system 102.
- the comparison is performed by the forecasting server 104.
- the comparison may result in one or more matching vector sequences 206.
- the matching vector sequences 206 have the same or similar data from past observations.
- the monitored system is a city and the target information is the pollen count
- the observed vector sequence 202 may contain meteorological information, such as temperature information, barometric information, and rainfall records for the last 30 days.
- Each matching vector sequence 206 may be associated with a day or time period that has similar temperature information, barometric information and rainfall records as for last 30 days and a pollen count may be obtained for the days associated with each matching vector sequence 206. It is understood that the comparison may also result in zero matching vector sequences found.
- FIG. 3 illustrates a second step of the enhanced SKR modeling when additional posterior information is used.
- Each matching vector sequence 206 is a snap shot of the behavior of the monitored system in a specific time period and posterior information is the additional data for the monitored system taken after this snap shot period.
- the forecasting, using the enhanced SKR modeling according to the embodiments, is improved when this additional data (posterior information) are used.
- a user 110 may specify, by using the user device 108, additional future parameters to be considered and these additional parameters are formed into a future vector sequence 302.
- the forecasting server 104 compares this future vector sequence 302 with posterior vector sequences 304 associated with the matching vector sequence 206. These posterior vector sequences 304 contain additional data that are future data in relation to the matching vector sequence 206. The comparison may result in one or more resulting vector sequences 306. From these resulting vector sequences, the user may derive desired information. The user may change the parameters in the future vector sequence 302 and then obtain a different set of resulting vector sequences 306. The user may also repeat this second step shown in FIG. 3 to further reduce the number of the resulting vector sequences 306.
- the comparison steps described in FIGs. 2 and 3 are performed using traditional SKR modeling techniques.
- these two sets of vector sequences can be merged into one single set.
- the merger can be done by assigning different weights for vector sequences from different sets, by simple combination, or other suitable methods.
- the merged set of vector sequences will then be used by an enhanced SKR based model to generate a set of estimate vector sequences.
- the generation of the estimate vector sequences using the enhanced SKR based model has been well described in the US Patent App. Nos. 13/186,153, the specification of which is incorporated by reference, and will not be repeated here.
- the user may want to know the pollen count for the city when certain conditions occur. For example, what if the expected rainfall increases and the average wind speed also increases over the next 30 days, how would the pollen count be affected? The user may add these rainfall and wind conditions in the future vector sequence 302 and this future vector sequence 302 is compared with the future conditions for the matching vector sequences 206.
- the forecasting server 104 will produce one or more resulting vector sequences 306 and each resulting vector sequence 306 has a pollen count for the future day. From these resulting vector sequences 306, a better forecast can be made about the pollen count for the future when certain conditions are predicted for the future. It is noted that future vector sequence 302 may contain parameters previously not considered in the first comparison and additional refinement illustrated in FIG. 3 may be made.
- FIG. 4 is an exemplary flowchart of the forecast method according to the embodiments.
- a data vector sequence (observed vector sequence) is received from a monitored system, step 402, by the forecasting server 104 and the data vector sequence is compared with stored data vector sequence, step 404.
- a set of matching vector sequences is obtained from this comparison, step 406.
- the user may enter a set of future conditions and this forecast vector sequence (future vector sequence) is received from the user (step 408).
- This set of future conditions is compared with posterior data associated with each matching vector sequence, step 410, and a new set of resulting vector sequence is generated.
- This set of resulting vector sequences and the set of matching vector sequences are merged to generate a set of merged vector sequences, step 412.
- This set of merged vector sequences is then used by an enhanced SK model to generate a set of estimate vector sequences, step 414.
- the estimate vector sequences are compared with the future vector sequence input by the user to determine if there were any problem in calculation, step 416, and the result is displayed to the user, step 418. An alert is generated if a difference larger than an expected threshold is detected.
- FIG. 6 illustrates an exemplary architecture 600 of a forecasting server 104 according to the embodiments.
- the present invention is not limited to the architecture 600 depicted in FIG. 6.
- the forecasting server 104 has a monitoring port interface 604 for receiving monitored data (observed vector sequence data) from a monitored system 102, along with a database interface 606 for retrieving stored data (stored data vector sequence data) from a database server 106.
- a vector sequence comparator 608 is provided for formulating, comparing, and analyzing monitored data received from the monitored system 102 against the retrieved data from the database server 106.
- a user interface 610 is provided for receiving additional future data (future vector sequence data), and a storage unit 610 stores a control program (not shown) that controls the vector sequence comparator 608.
- the implementation of the enhanced SK model according to the embodiments enables more accurate forecast for a monitored system.
- an operator may want to know the power that the wind turbine can produce if certain environment conditions were changed.
- the operator can enter environmental information, such as wind speed and the variation of the wind speed for the next few days. Consequently, the forecasting server will use this future wind speed information and data, provided by the sensors in the monitored wind turbine, to produce a more accurate forecast of the power produced by the wind turbine for these next few days.
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- General Physics & Mathematics (AREA)
- Automation & Control Theory (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A method using posterior information for improving forecasting capabilities is disclosed. The method implements an enhanced Sequential Kernel Regression (SKR) model by using posterior information associated with each observed conditions. The method generates a set of observed conditions for a monitored system and compares the posterior information for these conditions with a set of future parameters input by a user. From this comparison, an enlarged set of vector sequences is generated and used by the enhanced SKR model to generate a set of estimated vector sequences.
Description
METHOD FOR USING POSTERIOR INFORMATION TO IMPROVE FORECAST CAPABILITIES OF SEQUENTIAL KERNEL REGRESSION MODELS
FIELD OF THE INVENTION
[0001] The present invention generally relates to kernel regression modeling, and more specifically to use of the kernel regression modeling to forecast future behavior of a monitored system.
BACKGROUND OF THE INVENTION
[0002] Conventional Sequential Kernel Regression (SKR) is an empirical modeling method that merges the powerful multivariate pattern recognition capabilities of kernel regression with explicit pattern matching through time to provide unique capabilities for forecasting multiple sensor trends across wide time scales. SKR utilizes a pattern matching technique that matches an input pattern sequence, composed of a sequence of recent observation vectors ending with the current observation, to arrays of contiguous observation vectors in the training matrix to identify periods of past operation most similar to current system behavior.
[0003] Each of the identified periods of past operation has its own future, represented by observation vectors that succeed the matched vectors. Operating on the data sequences from the training matrix that are the best matches to current system behavior, the modeling technique produces forecasts of future system behavior. In this manner, the modeling technique utilizes prior information, as represented by the input pattern sequence, to produce forecasts of future behavior.
[0004] Conventional SKR based modeling provides forecasts of possible system behavior utilizing prior information only. These forecasts do not account for expectations or constraints on future system behavior though.
SUMMARY OF THE EMBODIMENTS
[0005] What is needed, therefore, are methods and systems to improve the ability of SKR based modeling to deliver refined forecasts based on independently-derived expectations or constraints on future conditions.
[0006] Embodiments of the present invention overcome the aforementioned deficiencies noted in the conventional modeling methods. An aspect of embodiments of the present invention is to provide a method for improving future forecast by using known posterior information for a set of candidate conditions.
[0007] In one embodiment, a method for forecasting future behaviors of a system is provided. The method includes receiving, by a forecasting server, a plurality of data, forming a sequence of observed vectors with the plurality of data, comparing the observed sequence of vectors with a plurality of stored vector sequences. The method also includes generating a plurality of matching vectors according to a comparison result, comparing a plurality of future conditions with posterior information associated with each matching vector. A sequence of forecast vectors is obtained with future behaviors based on comparison of the plurality of future conditions with the posterior information from each matching vector.
[0008] In an alternate embodiment, an apparatus is provided for improving forecast of future behavior of a system. The apparatus includes a monitoring port interface for receiving monitoring data from a monitored system and a vector sequence comparator for forming a sequence of observed vectors with the monitoring data. The observed vector sequence is compared with a plurality of stored vector sequences. The apparatus also includes generating a plurality of matching vectors based on a comparison result, comparing the plurality of matching vectors with a future condition, and generating a sequence of forecast vectors based on comparison between the plurality of matching vectors with the future condition.
[0009] In yet another alternate embodiment, a non-transient computer-readable storage medium is provided having computer executable instructions, when executed by a computer, causes the computer to perform certain operations. The operations include receiving a plurality of data, forming a sequence of observed vectors with the plurality of data, comparing the observed vector sequence with a plurality of stored vector sequences.
The operation also includes generating a plurality of matching vectors according to a comparison result, and comparing a plurality of future conditions with posterior information associated with each matching vector. A sequence of forecast vectors is obtained with future behaviors based on comparison of the plurality of future conditions with the posterior information from each matching vector.
[0010] The foregoing and other objects, features, aspects and advantages of the present invention will become better understood from a careful reading of a detailed description provided herein below with appropriate reference to the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
[0011] The present invention can be understood in more detail by reading the subsequent detailed description in conjunction with the examples and references made to the accompanying drawings, wherein:
[0012] FIG. 1 shows architecture of a forecast system according to the present invention;
[0013] FIG. 2 illustrates a first step in the forecast method according to the invention;
[0014] FIG. 3 illustrates a second step 300 in the forecast method using posterior information according to the present invention;
[0015] FIG. 4 is a flowchart of the forecast method according to the present invention;
[0016] FIG. 5 is a chart illustrating comparative forecasting results; and
[0017] FIG. 6 illustrates architecture of a forecasting device according to the present invention.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0018] While the present invention is described herein with illustrative embodiments for particular applications, it should be understood that the invention is not limited thereto. Those skilled in the art with access to the teachings provided herein will recognize additional modifications, applications, and embodiments within the scope thereof and additional fields in which the invention would be of significant utility.
[0019] The embodiments improve upon the forecasting and prognostics capabilities of the inventions described in US Patent App. Nos. 13/186,153 and App. No. 13/186,200, the specification of which are incorporated here in their entirety by this reference.
[0020] FIG. 1 is architecture 100 of a forecasting and prognostic system according to the invention. A device 102 is a device to be monitored and is set with many sensors for collecting the data at each sensing point. The collected data are sent to the forecasting server 104. The forecasting server 104 compares the collected data with data from a database 106 and produces a forecast that is displayed on a user device 108 to the user 110. The forecasting server 104 compares and produces the forecast by using Sequential Kernel Regression Modeling technique.
[0021] By way of background, enhanced SKR modeling, as provided via embodiments of the present invention, significantly improves forecasting of modeled parameters over the conventional SKR modeling technologies. Both enhanced and conventional SKR merge the powerful cross-parameter pattern recognition capabilities of kernel regression with explicit pattern matching through time to provide unique capabilities for extrapolating sensor trends across wide time scales.
[0022] Conventional SKR utilizes a pattern matching technique that matches the sequence of recent observation vectors ending with the current observation to equal-sized sequences from past operation of the modeled system. The most recent sequence of vectors, called the input pattern sequence, is an array containing data from the set of model variables collected during a contiguous set of time points ending with the most recent time point. A set of observation vectors that precede the current observation vector is referred to herein as the look back window. The number of vectors in it is called the look back count.
[0023] The conventional SKR compares the input pattern sequence to arrays of contiguous observation vectors in the training matrix to identify periods of past operation most similar to current system behavior. Each of the periods in the past identified by the pattern matching process has their own future, represented by observation vectors from time points that succeed the matched vectors. The conventional SKR technique collects a set of succeeding vectors, referred to herein as the look ahead window, from each of the time periods it has matched to the input pattern sequence. The SKR modeling
mathematical expression operates on the data in the input pattern sequence and the data in the matched periods from the past to calculate an array of estimate vectors. Like conventional SKR, enhanced SKR utilizes pattern matching to identify periods in the past that are similar to data in the input pattern sequence. But where enhanced SKR departs from conventional SKE, it is that the pattern matching algorithm is augmented by expectations of future system response to identify periods in the past that are similar to the input pattern sequence and the expected future behavior.
[0024] Each estimate vector contains an estimate of the expected behavior of all modeled variables at a point in time. The vectors in the estimate array correspond to the time points in the look back window, the current time point, and the time points in the look ahead window. The data in the look ahead window are estimates of the expected future behavior of the modeled variables. Thus, the enhanced SKR technique, in accordance with the embodiments, produces temporal sequences (trends) for all modeled variables, spanning time from the recent past, through the current time, and into the future.
[0025] More specifically, enhanced SKR incorporates posterior information to improve upon the pattern matching process of conventional SKR. As understood by those of skill in the art, posterior information is data about future events that has been produced by means independent of those employed by the sequential kernel regression model. By way of example, consider the use of aspects of the embodiments to improve the performance of SKR models of wind turbine site behavior using meteorological wind speed forecast data. An SKR model of a wind turbine includes data from sensors that measure site wind speed and measurements of the power produced by individual and/or groups of turbines.
[0026] Another example of posterior information is econometric forecasts of market conditions for use with SKR models of financial time series data, such as stock indexes, equity prices, and interest rates. For financial applications, the embodiments provide a new means to perform scenario analysis - the process of analyzing the effect of possible future events on the behavior of financial instruments.
[0027] Yet another example of posterior information is the end state of an equipment fault for use with SKR models of industrial equipment. In this application, the
embodiments provide a new means of performing prognostics by estimating the probability of various fault scenarios and remaining useful life of the equipment given current evidence of equipment degradation.
[0028] In an exemplary illustration of performance of the embodiments, an SKR model of a group of wind turbines co-located at a single geographic site was modified to use meteorological wind speed forecast data. Sensor data from a site containing 52 wind turbines were analyzed with an SKR model. Sensor data included in the model consisted of the power produced by 8 turbine groups, the total site power, and a single site wind speed signal. These data were augmented with a set of data files that contained 48 hour forecasts of various meteorological parameters, including wind speed and direction for the site.
[0029] A meteorological forecast file was provided for each hour of wind site operation over a 17 month period. The wind speed data were extracted from the forecast files, providing a rolling 48 hour forecast of wind conditions. Wind speed forecast data were not directly included in the model. Instead, the embodiments operate on the forecast wind speed data to improve pattern matching between current and past periods of operation. As an aspect of the embodiments, two tests were used to improve pattern matching.
[0030] In a first test, the standard pattern matching between the input pattern sequence and sequences of observation vectors in the training matrix was augmented by pattern matching of the forecast wind speed to wind speed data from time points in the training matrix that succeed vectors matched by the standard algorithm. This was accomplished by performing two similarity calculations.
[0031] In the first calculation, the similarity between the input pattern sequence and training matrix was calculated. In the second calculation, the similarity between the forecast wind speed data and wind speed data from succeeding vectors in the training matrix was calculated. The average value of the two similarities is then calculated. Those regions in the training matrix that produce the largest average similarity values are identified by the pattern matching technique and used by the SKR modeling equation to calculate an array of estimate vectors.
[0032] In a second test, pattern matching of the forecast wind speed to wind speed
data in the training matrix is the only method used to identify periods of past operation most similar to current system behavior.
[0033] Results from the two tests, noted above, are captured in FIG. 5. More specifically, FIG. 5 is an illustration of standard error for the forecast of a total power signal as a function of a look ahead count. Results from three calculations are illustrated in FIG. 5. Results from calculations using the two tests, aspects of the embodiments, are depicted by lines 502 and 504. Results obtained using the conventional SKR algorithm are depicted by line 506. Study of FIG. 5 reveals that by using meteorological forecast data to improve pattern matching, in accordance with the embodiments, the long-term ability of the SKR models to forecast site behavior is substantially improved.
[0034] In the embodiments, the conventional SKR pattern matching between an input pattern sequence and sequences of observation vectors in a training matrix is augmented. This augmentation occurs by pattern matching of the forecast wind speed to wind speed data from observation vectors in the training matrix that succeed vectors matched by the conventional SKR. The enhanced SKR improves pattern matching, ultimately providing more accurate forecasts of site -wide turbine power production.
[0035] FIG. 2 is an illustration 200 of a first step in the enhanced SKR modeling according to embodiments of the present invention. Data from recent observations of the monitored system 102 is put into a vector sequence format. The observed vector sequence 202 of the recent observations is compared with a plurality of past vector sequences (a.k.a. stored vector sequences) 204 retrieved from the database 106. The vector sequences retrieved from the database 106 are data collected from past observations of the monitored system 102.
[0036] In FIG. 2, the comparison is performed by the forecasting server 104. The comparison may result in one or more matching vector sequences 206. The matching vector sequences 206 have the same or similar data from past observations. As an example, if the monitored system is a city and the target information is the pollen count, the observed vector sequence 202 may contain meteorological information, such as temperature information, barometric information, and rainfall records for the last 30 days.
[0037] Each matching vector sequence 206 may be associated with a day or time period that has similar temperature information, barometric information and rainfall
records as for last 30 days and a pollen count may be obtained for the days associated with each matching vector sequence 206. It is understood that the comparison may also result in zero matching vector sequences found.
[0038] FIG. 3 illustrates a second step of the enhanced SKR modeling when additional posterior information is used. Each matching vector sequence 206 is a snap shot of the behavior of the monitored system in a specific time period and posterior information is the additional data for the monitored system taken after this snap shot period. The forecasting, using the enhanced SKR modeling according to the embodiments, is improved when this additional data (posterior information) are used.
[0039] A user 110 may specify, by using the user device 108, additional future parameters to be considered and these additional parameters are formed into a future vector sequence 302. The forecasting server 104 compares this future vector sequence 302 with posterior vector sequences 304 associated with the matching vector sequence 206. These posterior vector sequences 304 contain additional data that are future data in relation to the matching vector sequence 206. The comparison may result in one or more resulting vector sequences 306. From these resulting vector sequences, the user may derive desired information. The user may change the parameters in the future vector sequence 302 and then obtain a different set of resulting vector sequences 306. The user may also repeat this second step shown in FIG. 3 to further reduce the number of the resulting vector sequences 306. The comparison steps described in FIGs. 2 and 3 are performed using traditional SKR modeling techniques.
[0040] After obtaining two sets of vector sequences, one from the comparison described in FIG. 2 and one from the comparison described in FIG. 3, these two sets of vector sequences can be merged into one single set. The merger can be done by assigning different weights for vector sequences from different sets, by simple combination, or other suitable methods. The merged set of vector sequences will then be used by an enhanced SKR based model to generate a set of estimate vector sequences. The generation of the estimate vector sequences using the enhanced SKR based model has been well described in the US Patent App. Nos. 13/186,153, the specification of which is incorporated by reference, and will not be repeated here.
[0041] For the example of the pollen count described above, the user may want to
know the pollen count for the city when certain conditions occur. For example, what if the expected rainfall increases and the average wind speed also increases over the next 30 days, how would the pollen count be affected? The user may add these rainfall and wind conditions in the future vector sequence 302 and this future vector sequence 302 is compared with the future conditions for the matching vector sequences 206.
[0042] After the comparison, the forecasting server 104 will produce one or more resulting vector sequences 306 and each resulting vector sequence 306 has a pollen count for the future day. From these resulting vector sequences 306, a better forecast can be made about the pollen count for the future when certain conditions are predicted for the future. It is noted that future vector sequence 302 may contain parameters previously not considered in the first comparison and additional refinement illustrated in FIG. 3 may be made.
[0043] FIG. 4 is an exemplary flowchart of the forecast method according to the embodiments. In FIG. 4, a data vector sequence (observed vector sequence) is received from a monitored system, step 402, by the forecasting server 104 and the data vector sequence is compared with stored data vector sequence, step 404. A set of matching vector sequences is obtained from this comparison, step 406. The user may enter a set of future conditions and this forecast vector sequence (future vector sequence) is received from the user (step 408). This set of future conditions is compared with posterior data associated with each matching vector sequence, step 410, and a new set of resulting vector sequence is generated. This set of resulting vector sequences and the set of matching vector sequences are merged to generate a set of merged vector sequences, step 412. This set of merged vector sequences is then used by an enhanced SK model to generate a set of estimate vector sequences, step 414. The estimate vector sequences are compared with the future vector sequence input by the user to determine if there were any problem in calculation, step 416, and the result is displayed to the user, step 418. An alert is generated if a difference larger than an expected threshold is detected.
[0044] FIG. 6 illustrates an exemplary architecture 600 of a forecasting server 104 according to the embodiments. However, the present invention is not limited to the architecture 600 depicted in FIG. 6.
[0045] In FIG. 6, the forecasting server 104 has a monitoring port interface 604 for
receiving monitored data (observed vector sequence data) from a monitored system 102, along with a database interface 606 for retrieving stored data (stored data vector sequence data) from a database server 106. A vector sequence comparator 608 is provided for formulating, comparing, and analyzing monitored data received from the monitored system 102 against the retrieved data from the database server 106. A user interface 610 is provided for receiving additional future data (future vector sequence data), and a storage unit 610 stores a control program (not shown) that controls the vector sequence comparator 608.
[0046] The implementation of the enhanced SK model according to the embodiments enables more accurate forecast for a monitored system. In the monitored wind turbine site example above, an operator may want to know the power that the wind turbine can produce if certain environment conditions were changed. In this scenario, the operator can enter environmental information, such as wind speed and the variation of the wind speed for the next few days. Consequently, the forecasting server will use this future wind speed information and data, provided by the sensors in the monitored wind turbine, to produce a more accurate forecast of the power produced by the wind turbine for these next few days.
[0047] Although the present invention has been described with reference to the preferred embodiments, it will be understood that the invention is not limited to the details described thereof. Various substitutions and modifications have been suggested in the foregoing description, and others will occur to those of ordinary skill in the art. Therefore, all such substitutions and modifications are intended to be embraced within the scope of the invention as defined in the appended claims. It is understood that features shown in different figures can be easily combined within the scope of the invention.
Claims
CLAIMS OF THE INVENTION
What is claimed is:
1. A method, for forecasting future behaviors of a system, comprising:
forming an observed vector sequence with a plurality of data received via a forecasting server;
generating a plurality of matching vector sequences according to a comparison between the observed vector sequence with a plurality of stored vectors;
comparing a plurality of future conditions with posterior information associated with each matching vector sequence;
obtaining a plurality of resulting vector sequences with future behaviors based on a comparison result;
merging the plurality of the matching vector sequences with the plurality of resulting vector sequences to form a merged vector sequences; and
generating, by an enhanced SKR model based on the merged vector sequences, a set of estimate vector sequences.
2 The method of claim 1 , wherein the obtaining is based on comparison of the plurality of future conditions with the posterior information from each matching vector.
3. The method of claim 1 further comprising the step of collecting the plurality of data from a plurality of monitoring points in the system.
4. The method of claim 1 further comprising the steps of receiving, by the forecasting server, a plurality of future conditions from a user input device and forming a future vector sequence.
5. The method of claim 4, further comprising the steps of:
checking a difference between the set of estimate vector sequences and the future vector sequence; and
generating an alert if the difference is larger than a threshold.
6. The method of claim 1 further comprising the step of retrieving the plurality of stored vectors from a database, where each stored vector containing past observation data.
7. The method of claim 1, wherein the comparing steps are performed using a Sequential Kernel Regression algorithm.
8. An apparatus, for improving forecast of future behavior of a system, comprising: a monitoring port interface for receiving monitoring data from a monitored system; and
a vector sequence comparator for
forming an observed vector sequence with the monitoring data, comparing the observed vector sequence with a plurality of stored vectors, generating a plurality of matching vector sequences based on a
comparison result,
comparing a plurality of future conditions with posterior information associated with each matching vector sequence,
generating a plurality of resulting vector sequences based on comparison between the plurality of matching vector sequences with the plurality of future conditions,
merging the plurality of resulting vector sequences with the plurality of matching vector sequences to form a plurality of merged vector sequences, and generating a plurality of estimate vector sequences .
9. The system of claim 8, further comprising a database interface for retrieving the plurality of stored vectors from a database.
10. The apparatus of claim 8, further comprising a user interface for receiving a future condition from a user to form a future vector sequence.
11. The apparatus of 10, wherein the vector sequence comparator further performs
checking a difference between the plurality of estimate vector sequences and the future vector sequence; and
generating an alert if the difference is larger than a threshold.
12. The apparatus of claim 8, further comprising a storage unit for storing a control program for controlling vector sequence comparator.
13. The apparatus of claim 12, wherein the control program uses a Sequential Kernel Regression algorithm.
14. A non-transient computer-readable storage medium having computer executable instructions, when executed by a computer, causes the computer to perform operations comprising;
receiving a plurality of data;
forming an observed vector sequence with the plurality of data;
comparing the observed vector sequence with a plurality of stored vectors;
generating a plurality of matching vector sequences according to a comparison result;
comparing a plurality of future conditions with posterior information associated with each matching vector sequence;
obtaining a plurality of resulting vector sequences with future behaviors based on a comparison result;
merging the plurality of the matching vector sequences with the plurality of resulting vector sequences to form a merged vector sequences; and
generating, by an enhanced SKR model based on the merged vector sequences, a set of estimate vector sequences.
16. The computer executable instructions of claim 15, wherein the obtaining is based on comparison of the plurality of future conditions with the posterior information from each matching vector sequence.
17. The computer executable instructions of claim 15 further causing the computer to perform an operation for collecting the plurality of data from a plurality of monitoring points in the system.
18. The computer executable instructions of claim 15 further causing the computer to perform an operation for receiving a plurality of future conditions from a user input device and forming a future vector sequence.
19. The computer executable instructions of claim 15 further causing the computer to perform an operation for retrieving the plurality of stored vectors from a database, where each stored vector containing past observation data.
20. The computer executable instructions of claim 15, wherein the comparing steps are performed using a Sequential Kernel Regression algorithm
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/US2013/039456 WO2014178877A1 (en) | 2013-05-03 | 2013-05-03 | Method for using posterior information to improve forecast capabilities of sequential kernel regression models |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/US2013/039456 WO2014178877A1 (en) | 2013-05-03 | 2013-05-03 | Method for using posterior information to improve forecast capabilities of sequential kernel regression models |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2014178877A1 true WO2014178877A1 (en) | 2014-11-06 |
Family
ID=48468792
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/US2013/039456 WO2014178877A1 (en) | 2013-05-03 | 2013-05-03 | Method for using posterior information to improve forecast capabilities of sequential kernel regression models |
Country Status (1)
Country | Link |
---|---|
WO (1) | WO2014178877A1 (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108885718A (en) * | 2016-01-14 | 2018-11-23 | 摄取技术有限公司 | Localize time model prediction |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6216119B1 (en) * | 1997-11-19 | 2001-04-10 | Netuitive, Inc. | Multi-kernel neural network concurrent learning, monitoring, and forecasting system |
US20130024414A1 (en) * | 2011-07-19 | 2013-01-24 | Smartsignal Corporation | System of Sequential Kernel Regression Modeling for Forecasting and Prognostics |
-
2013
- 2013-05-03 WO PCT/US2013/039456 patent/WO2014178877A1/en active Application Filing
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6216119B1 (en) * | 1997-11-19 | 2001-04-10 | Netuitive, Inc. | Multi-kernel neural network concurrent learning, monitoring, and forecasting system |
US20130024414A1 (en) * | 2011-07-19 | 2013-01-24 | Smartsignal Corporation | System of Sequential Kernel Regression Modeling for Forecasting and Prognostics |
Non-Patent Citations (3)
Title |
---|
JUBAN J ET AL: "Probabilistic Short-term Wind Power Forecasting for the Optimal Management of Wind Generation", POWER TECH, 2007 IEEE LAUSANNE, IEEE, PISCATAWAY, NJ, USA, 1 July 2007 (2007-07-01), pages 683 - 688, XP031269459, ISBN: 978-1-4244-2189-3 * |
MIGUEL G LOBO ET AL: "Regional Wind Power Forecasting Based on Smoothing Techniques, With Application to the Spanish Peninsular System", IEEE TRANSACTIONS ON POWER SYSTEMS, IEEE SERVICE CENTER, PISCATAWAY, NJ, US, vol. 27, no. 4, 1 November 2012 (2012-11-01), pages 1990 - 1997, XP011470004, ISSN: 0885-8950, DOI: 10.1109/TPWRS.2012.2189418 * |
RICARDO J BESSA ET AL: "Time Adaptive Conditional Kernel Density Estimation for Wind Power Forecasting", IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, IEEE, USA, vol. 3, no. 4, 1 October 2012 (2012-10-01), pages 660 - 669, XP011462107, ISSN: 1949-3029, DOI: 10.1109/TSTE.2012.2200302 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108885718A (en) * | 2016-01-14 | 2018-11-23 | 摄取技术有限公司 | Localize time model prediction |
EP3403220A4 (en) * | 2016-01-14 | 2019-09-11 | Uptake Technologies, Inc. | Localized temporal model forecasting |
AU2017206794B2 (en) * | 2016-01-14 | 2022-01-20 | Uptake Technologies, Inc. | Localized temporal model forecasting |
US11295217B2 (en) | 2016-01-14 | 2022-04-05 | Uptake Technologies, Inc. | Localized temporal model forecasting |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US20150254554A1 (en) | Information processing device and learning method | |
EP3112959B1 (en) | Method for detecting anomalies in a water distribution system | |
JP6243080B1 (en) | Preprocessor and abnormal sign diagnosis system | |
US10496730B2 (en) | Factor analysis device, factor analysis method, and factor analysis program | |
Kansal et al. | Detection of forest fires using machine learning technique: A perspective | |
Richman et al. | Missing data imputation through machine learning algorithms | |
JP2017194341A (en) | Abnormality diagnosis method, abnormality diagnosis device, and abnormality diagnosis program | |
US11216534B2 (en) | Apparatus, system, and method of covariance estimation based on data missing rate for information processing | |
CN106598822B (en) | A kind of abnormal deviation data examination method and device for Capacity Assessment | |
US11416007B2 (en) | Computer-implemented method and system for evaluating uncertainty in trajectory prediction | |
KR20210017342A (en) | Time series prediction method and apparatus based on past prediction data | |
CN114297036A (en) | Data processing method and device, electronic equipment and readable storage medium | |
US20140188777A1 (en) | Methods and systems for identifying a precursor to a failure of a component in a physical system | |
Biswas et al. | Weather prediction by recurrent neural network dynamics | |
Oo et al. | Time series prediction based on Facebook Prophet: a case study, temperature forecasting in Myintkyina | |
JP7067234B2 (en) | Data discrimination program, data discrimination device and data discrimination method | |
US20170286841A1 (en) | Monitoring device and monitoring method thereof, monitoring system, and recording medium in which computer program is stored | |
Wang et al. | Reconstruction of missing trajectory data: a deep learning approach | |
JP2019105871A (en) | Abnormality candidate extraction program, abnormality candidate extraction method and abnormality candidate extraction apparatus | |
CN113110961B (en) | Equipment abnormality detection method and device, computer equipment and readable storage medium | |
JP6494258B2 (en) | Prediction system, prediction method, and prediction program | |
CN107924182A (en) | The control method of monitoring arrangement and monitoring arrangement | |
CN117540336A (en) | Time sequence prediction method and device and electronic equipment | |
CN117170915A (en) | Data center equipment fault prediction method and device and computer equipment | |
CN110874601B (en) | Method for identifying running state of equipment, state identification model training method and device |
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: 13724075 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 13724075 Country of ref document: EP Kind code of ref document: A1 |