IT201600125128A1 - METHOD FOR DETERMINING AN ENERGY MODEL OF A PLANT AND CORRESPONDING METHOD TO MONITOR THE ENERGY BEHAVIOR OF THE PLANT - Google Patents

Description

DESCRIPTION

of the patent for industrial invention entitled:

"METHOD FOR DETERMINING AN ENERGY MODEL OF A PLANT AND CORRESPONDING METHOD FOR MONITORING THE ENERGY BEHAVIOR OF THE PLANT"

by SEASIDE S.R.L.

of Italian nationality

with registered office: VIA ANDREA COSTA, 165

BOLOGNA (BO)

Inventors: BENFENATI Paolo, TEXTS Piergiorgio

* * *

The present invention relates to a method for determining an energy model of a plant, ie a complex system of apparatuses that consume and / or produce energy, and to a corresponding method for monitoring its energy behavior.

The exact determination of the correct energy behavior of a plant, for example a plant or an entire plant in the industrial sector or in the tertiary sector, is difficult since the variables that in principle influence it are very many, often not known or difficult to measure, and practically impossible to relate to each other.

Modern industrial plants are equipped with energy monitoring systems including a variety of

1

Andrea FIORINI (Registered on the Register N.1197 / B) devices designed to measure a plurality of independent variables that presumably affect the energy consumption of the system. These independent variables include, by way of example, physical quantities measured by environmental sensors, such as for example temperature, pressure and solar radiation sensors, and variables measured by meters and / or meters of various kinds, such as for example piece counters associated with automatic machines. , calorie counters, gas meters, flow meters and process step meters to measure quantities of raw materials and items.

Energy monitoring systems produce a considerable amount of data that is stored in one or more databases to be analyzed by a professional figure known as "energy manager". The most sophisticated energy monitoring systems provide the measured values in terms of time series. The amount of data contained in the databases grows exponentially with the increasing complexity (number of devices and devices) of the industrial plant.

In order to measure the energy performance of an industrial plant and evaluate its variations, the energy manager must define the so-called reference consumption (or "energy baseline") and then periodically measure any progress or regression with respect to this consumption of

2

Andrea FIORINI (Registration No.1197 / B) reference. Typically, the energy manager adopts as reference consumption a series of values calculated with very simple methods, usually based on average value and linear regression, then manually identifies, by trial and error, the relationship existing between independent variables and dependent variables and compares the energy value measured with reference consumption. Basically, the energy manager performs simple calculations with office automation tools on manually extracted data, often extemporaneously and with usually high granularity. This calculation method is unlikely to continuously optimize the energy performance of the system.

The object of the present invention is to provide a method for analyzing the energy behavior of an industrial plant, which is free from the drawbacks described above and, at the same time, is easy and economical to implement.

In accordance with the present invention, a method is provided for determining an energy model of a plant, a method for monitoring the energy behavior of a plant and a monitoring system of a plant, as defined in the attached claims.

The present invention will now be described with reference to the accompanying drawings, which illustrate an example thereof

3

Andrea FIORINI (Registration No.1197 / B) of non-limiting implementation, in which:

- Figure 1 illustrates a general architecture of a monitoring system for a plant, which system is configured to implement the method for determining an energy model of the plant and the method for monitoring the energy behavior of the plant made according to the present invention; And

- Figure 2 illustrates a flow chart of the method for determining an energy model of the plant.

In figure 1, with 1 a plant is generically indicated, i.e. a complex system of devices that consume and / or produce energy, such as an industrial plant, and with 2 a monitoring system of plant 1. The monitoring system 2 it comprises an essentially hardware subsystem, indicated with 3, and an essentially software subsystem, indicated with 4.

The hardware subsystem 3 comprises a measurement acquisition system 5 for acquiring measured values of a plurality of environmental and / or plant variables that affect the energy behavior or represent the energy consumption or production of the plant 1, and a transmission system data 6 consisting, for example, of a gateway of a communications network which receives the measured values from the measurement acquisition system 5 and transmits them outside the hardware subsystem 3.

4

Andrea FIORINI (Registered in the Albo N.1197 / B) Typically, the measured values of the variables are provided in the form of relative historical data series, indicated with HD1 in figure 1, i.e. at least one historical series of measured values for each measured variable. Preferably, the historical series of HD1 data are collected in text files in cvs format.

The measurement acquisition system 5 comprises a plurality of sensors and / or measurement devices, known per se and therefore not illustrated in detail, such as for example sensors for measuring physical quantities - for example temperatures, pressures of fluids and solar radiation - calorie counters, gas meters, electric meters, flow meters of fluids, counters of automatic machines to count volumes or pieces produced, and IoT devices, i.e. devices capable of communicating with the Internet according to the concept of the Internet of Things .

In the example of Figure 1, the hardware subsystem 3 also comprises a monitoring database 7 configured ad hoc to query the measurement acquisition system 5 and store the measured values to make them available at a later time already in the form of historical series of HD2 data, preferably collected in text files in cvs format.

Typically, the hardware subsystem 3 is located in the same place as plant 1. Instead, the subsystem

5

Andrea FIORINI (Registered in the Register N.1197 / B) software 4 can be implemented in a server that can be remotely displaced with respect to plant 1. The communication between the hardware subsystem 3 and the software subsystem 4, as well as the remote accessibility to the software subsystem 4 takes place through the internet according to known telecommunications technologies.

The software subsystem 4 comprises a management information system 8, which provides transactional data, again in the form of a plurality of historical data series indicated with HD3 in figure 1, relating to variables of management processes that affect the energy behavior of the plant 1 , a processing module 9 to clean and normalize the time series of HD1, HD2 and HD3 data in order to obtain time series of HD4 data that are uniform to each other, and a database 10 to store and make available all the variables measured or quantified through the measurement acquisition system 5 and the management information system 8, indicated with VAR, and the historical series of HD4 data originating from the VAR variables.

The software subsystem 4 also comprises a monitoring module 11, which receives in real time, during the normal operation of the plant 1, VME measured values of one or more plant variables to be monitored, which are a subset of the VAR variables. , from the system of

6

Andrea FIORINI (Registered in the Register N.1197 / B) acquisition of measurements 5 and from the management information system 8, preferably through the database 10, where these VME measured values are stored, and is configured to compare the VME measured values with the relative values of VREF reference, and to generate related alarm signals when VME measured values are greater than the relative VREF reference values.

In particular, the system variables that can be monitored are those that express the energy consumption of the system, if for example the latter is a system aimed at the production of articles or the processing of materials, or the energy production of the plant, if for example the latter is aimed at the production of electrical or thermal energy. By way of example, in a plant for the production of wall tiles, the plant variable to be monitored could be the consumption of methane gas.

According to the invention, the software subsystem 4 comprises a man-machine interface module 12, for example a web interface accessible through the Internet, to allow a user to select portions of such historical data series HD4 and a configured processing module 13 to process historical HD4 data on the basis of the selections made by the user in order to determine an energy model of the plant 1. The energy model

7

Andrea FIORINI (Registered in the Albo N.1197 / B) is then used to calculate the VREF reference values in real time as a function of VM measured values of another subset of the VAR variables.

The aforementioned VM measured values are provided continuously during the normal operation of the plant 1 by the measurement acquisition system 5 and by the management information system 8, preferably through the database 10, where these measured VM values remain stored at least temporarily.

In particular, the processing module 13 is configured to implement the method illustrated by the flow diagram of Figure 2 and described below.

The user interacts with the software subsystem 4 through the man-machine interface 12 to select, among the plurality of VAR variables, the variables to be considered independent, indicated with VARind, and at least one variable to be considered as dependent, indicated with VARdip ( step 101 of Figure 2), and to define two sets of historical data, which hereinafter will be called HDtrain historical training data and HDtest historical test data, by selecting two respective time portions of the HD4 historical data series (step 102). Preferably, the time portion that defines HDtrain historical training data is much greater than that which defines HDtest historical test data for reasons that will be

8

Andrea FIORINI (Registration No.1197 / B) more clear in the rest of this document.

A plurality of artificial intelligence algorithms are trained with the HDtrain historical training data in order to determine, for each of the algorithms, a predictive model of the VARdip variable as a function of the VARind variables (step 103).

Artificial intelligence algorithms are of a known type and comprise one or more algorithms selected from a group of algorithms comprising:

- random forest;

- first order splines;

- second order splines;

- artificial neural network;

- artificial neural network of auto-encoder type;

- stacked denoising artificial neural network Each artificial intelligence algorithm generally comprises a real algorithm, a plurality of calculation nodes and a map of weights and coefficients associated with the calculation nodes, the value of the weights and coefficients being configured during training the algorithm. Therefore, the aforementioned predictive model consists of the relative algorithm and the configuration of said weights and coefficients obtained with training.

Advantageously, the instructions of the artificial intelligence algorithms and the related weight maps e

9

Andrea FIORINI (Registered in the Register N.1197 / B) coefficients is stored in an algorithm database (not shown) associated with the processing module 13. For example, if the processing module 13 is implemented on a server, that is, it is implemented by a particular software installed on that server, the algorithm database is stored in a memory of that server.

The time length of the HDtrain training historical data must be sufficient to contain at least one complete cycle of periodicity of the energy behavior of plant 1. For example, if plant 1 consists of an air conditioning system, the HDtrain historical training data must have a temporal extension of at least one year to understand both the periodicity of the day and night, and the seasonal periodicity.

The training phase of all artificial intelligence algorithms foresees, for each algorithm, to calculate the value of an IG significance index for each VARind variable expressing the degree of influence of the VARind variable itself on the VARdip variable value (step 104). In particular, the IG significance index is a purity index expressing the value of total decrease of impurities, i.e. of total increase of purity, of the calculation nodes. More specifically, the IG significance index consists of the so-called Gini Index.

10

Andrea FIORINI (Registration No.1197 / B) At this point, each of the trained algorithms is tested. First of all, each of the trained algorithms, i.e. the respective predictive model, is fed with historical HDtest test data to determine relative predicted VPR values of the VARdip variable (step 105). The time length of HDtest historical test data can be significantly shorter than that of HDtrain historical training data. For example, if system 1 consists of an air conditioning system, the historical HDtest test data can have a temporal extension of one or two months.

For each trained algorithm, a value of at least one statistical index is calculated as a function of the predicted VPR values and the measured values of the historical HDtest test data relating to the VARdip variable (step 106). In particular, the statistical index is such that it decreases as the difference between the predicted VPR values provided by a certain trained algorithm decreases and the corresponding measured values of the historical HDtest test data relating to the VARdip variable. For example, the statistical index consists of the root mean square error (RMSE) or the mean absolute error (MAE) between the predicted VPR values provided by a certain trained algorithm and the corresponding measured values of the historical HDtest test data relating to the VARdip variable.

11

Andrea FIORINI (Registered in the Register N.1197 / B) Among the trained algorithms, the best algorithm is selected based on the statistical index values of all the trained algorithms (step 107). In particular, considering a statistical index that decreases as the difference between the predicted VPR values and the corresponding measured values of the historical HDtest test data of the VARdip variable decreases, the best algorithm is the trained algorithm with the lowest statistical index value.

As previously mentioned, the training phase of the algorithms includes, for each of them, a phase of calculating the value of an IG significance index for each of the VARind variables. An independent variable VARind with a low value of the IG significance index is a variable that has little influence on the predicted values of the dependent variable VARdip.

Among the VARind variables, those variables are selected, indicated below with VARimp, which are most influential on the value of the VARdip variable, based on their IG significance index calculated for the best algorithm (step 108). In particular, the VARimp variables are those whose IG significance index is greater than a predetermined threshold value, for example equal to 10%. Advantageously, the threshold value for the IG index is configured by the user through the man-machine interface 12, for example at the same time as they are selected

12

Andrea FIORINI (Registration No.1197 / B) the variables VARind and VARdip and the historical data of HDtrain training and HDtest tests (step 102 of figure 2).

If the VARimp variables represent a subset of the VARind variables, the best algorithm previously selected (step 107 of Figure 2) is retrained with a subset of the historical HDtrain training data relating only to the VARimp variables (step 109). The best retrained algorithm on said subset of historical HDtrain training data defines the energy model sought for the plant. If, on the other hand, the VARimp variables coincide with the VARind variables, then the best algorithm does not need to be retrained and already defines the energy model sought.

According to a variant of the invention not illustrated, the IG significance index is calculated only for the best algorithm. This allows to reduce the calculation time of the energy model.

According to the present invention, the energy model is determined by selecting, as the VARdip variable, a plant variable to be monitored and is used to calculate in real time the reference value VREF of this plant variable as a function of the measured values VM of the VARimp variables only. , acquired by the measurement acquisition system 5 and by the management information system 8 during the normal operation of plant 1. The value of

13

Andrea FIORINI (Registration No.1197 / B) VREF reference of the plant variable to be monitored is the so-called "energy baseline" of the plant. As described above, the monitoring module 11 is configured to generate alarm signals when the VME measured values of the plant variable are greater than the reference value VREF.

The use of the energy model determined according to the method of the invention allows to adapt the monitoring of the plant 1 to the long-term behaviors of the plant 1 itself, since the energy model describes in the best possible way the relationships between the most independent variables. significant and dependent variables. Furthermore, the definition of the energy model leads to a reduction in the number of independent variables to be measured (VARimp) with a consequent significant reduction in the installation and maintenance costs of the devices of the measurement acquisition system 5 and, at the same time, an increase in the reliability of the Collected data. In addition, the energy model can be used to calculate the energy baseline of plants similar to the one for which the energy model was calculated.

From the above it is clear that the method according to the invention is applicable to any type of plant, be it an industrial plant or the tertiary sector.

14

Andrea FIORINI (Registration No.1197 / B)

Claims (11)

CLAIMS 1. Method for determining an energy model of a plant, comprising: - by means of measurement acquisition means (5, 8) associated with the plant, to acquire historical series of measured values (HD4) of a plurality of environmental and / or plant variables (VAR) that affect energy behavior or represent energy consumption o energy production of the plant (1); - selecting, from among said plurality of variables (VAR), first variables (VARind) to be considered as independent variables and at least a second variable (VARdip) to be considered as dependent variable; - define a first set of historical data (HDtrain) and a second set of historical data (HDtest) by selecting a first time portion and, respectively, a second time portion of said time series of measured values, preferably the second time portion being shorter of the first temporal portion; - train a plurality of artificial intelligence algorithms with the first set of historical data (HDtrain); - feeding each of the trained algorithms with the second set of historical data (HDtest) to determine relative predicted values (VPR) of said second variable (VARdip); 15 Andrea FIORINI (Registered in the Albo N.1197 / B) - for each trained algorithm calculate a value of at least one statistical index as a function of the predicted values (VPR) and the measured values of the second set of historical data (HDtest) relating to the second variable ( VARdip); - among the trained algorithms, select a first algorithm on the basis of the statistical index values of all the trained algorithms; - at least for the first algorithm, for each of the first variables (VARind), calculate a significance index (IG) expressing the degree of influence of the first variables (VARind) on the value of the second variable (VARdip); - among the first variables (VARind), select third variables (VARimp) as the most influential on the basis of their significance index (IG); - if the third variables (VARimp) are a subset of the first variables (VARind), then retrain the first algorithm with a subset of said first historical data set (HDtrain) related to said third variable (VARimp) and said energy model is defined by the first retrained algorithm, otherwise said energy model is defined by said first not retrained algorithm. Method according to claim 1, in which said historical series of measured values (HD4) are stored 16 Andrea FIORINI (Registered in the Register N.1197 / B) in a database (10) and said first variables (VARind) and said second variable (VARdip) are selected by a user through a man-machine interface (12) configured to access to said database (10). Method according to claim 1 or 2, wherein said historical series of measured values (HD4) are stored in a database (10) and said first and second time portions are selected by a user through a man-machine interface ( 12) configured to access said database (10). Method according to claim 1 to 3, wherein said statistical index being such that it decreases as the difference between said predicted values (VPR) provided by a certain trained algorithm and the corresponding measured values of the second set of historical data ( HDtest) related to the second variable (VARdip); said first algorithm being the trained algorithm with the lowest statistical index value. Method according to claim 1 to 3, wherein said statistical index is the mean square deviation or the mean absolute error between said predicted values (VPR) provided by a certain trained algorithm and the corresponding measured values of the second set of data historical (HDtest) relating to the second variable (VARdip); said first algorithm being the algorithm trained with the minor 17 Andrea FIORINI (Registration No.1197 / B) statistical index value. Method according to one of claims 1 to 5, wherein each artificial intelligence algorithm comprises a plurality of computing nodes; said significance index (IG) being a purity index expressing the total purity increase value of the calculation nodes. Method according to one of claims 1 to 5, wherein each artificial intelligence algorithm comprises a plurality of computing nodes; said significance index (IG) being the Gini Index. 8. Method according to one of claims 1 to 5, in which defining third variables (VARimp) by selecting them among the first variables (VARind) on the basis of their significance index (IG) comprises: - select as third variables (VARimp) those first variables (VARind) whose significance index (IG) is greater than a predetermined threshold value, for example equal to 10%. Method according to one of claims 1 to 8, wherein said artificial intelligence algorithms comprise one or more algorithms selected in a group of algorithms comprising: - random forest; - first order splines; - second order splines; 18 Andrea FIORINI (Registration No.1197 / B) - artificial neural network; - artificial neural network of auto-encoder type; - stacked denoising auto-encoder artificial neural network. 10. Method for monitoring the energy behavior of a plant, including: - through measurement acquisition means (5, 8) associated with the plant (1), acquire measured values (VME) of at least one plant variable that expresses the energy consumption or energy production of the plant (1); - comparing the measured values (VME) of said plant variable with a respective reference value (VREF); - generate an alarm signal when the measured values (VME) of the plant variable are greater than the reference value (VREF); and being characterized in that it comprises: - determining an energy model of the plant (1) in accordance with the method according to one of claims 1 to 9, in which said second variable (VARdip) coincides with said plant variable; - by means of said measurement acquisition means (5, 8), to acquire in real time measured values (VM) of said third variables (VARimp); And - calculate said reference value in real time 19 Andrea FIORINI (Registered in the Albo N.1197 / B) (VREF) using the energy model as a function of the measured values (VM) of the third variables (VARimp). 11. Monitoring system of a plant, comprising means for acquiring measurements (5, 8) associated with the plant (1) to acquire historical series of measured values (HD4) of a plurality of environmental and / or plant variables (VAR ), in real time during normal plant operation (1), a database (10) to store the historical series of measured values (HD4), processing means (13) interfaced with the database (10), and a human-machine interface (12) to allow a user to provide instructions to the processing means (13); said processing means (13) being configured to implement the method according to one of claims 1 to 10. p.i .: SEASIDE S.R.L. Andrea FIORINI 20 Andrea FIORINI (Registration No.1197 / B)