The FDS procedure
The kernel of the system is the neural networks. So, it is necessary to consider a first step to train the networks. This period should be as short as possible because during this period the system has
to be inspected regularly to be sure that nothing has happened (e. g. it is necessary to clean the collectors). It has been found that getting 30 values with a sampling period of 1 hour is a good compromise between sensitivity of the procedure and initial length.
At the end of the initialization period, three networks are available. They are used to estimate the temperatures. The latter are compared to the acquired temperatures through the computation of the RMSE. If the error is very small, the initial connection weights are stored in a matrix. If the error is higher than a threshold (0.05 for the collector array, 0.015 for the connecting pipes), the networks are re-trained. The new connection weights are stored in the connection weights matrix. If a given number of consecutive connection weights are different from the initial weights, an alarm is fired. It has been found that 5 consecutive values are sufficient, and do not lead to long delays between the fault and its detection.
 |
To generate the data, the typical meteorological year (TMY) files of Nicosia have been considered. Figure 4 shows the detection time for the F' drift considering the two draw off profiles.
Although it seems that the detection is quite late, a plot of the temperatures about the detection time shows that it would be very difficult to detect the drift by the simple analysis of the temperatures (Fig. 5).
|
|
|
Fig. 5 Differential between temperatures when the system is stable all year long and when F' evolves |
Figure 6 shows the detection time for the UL drift. On the one hand, it can be noted that a combined drift leads to an earlier detection. On the other hand, it has been checked that the defaults on the connecting pipes do not lead to any detection by the analysis of the collector array data.
 |
An on-line fault diagnostic system has been presented. The main advantage of this system is that it does not need a long training period. It has been shown that the drifts are detected well before the increase of the auxiliary electrical power is higher than 7.5%. This means that the FDS is sensitive. As the neural networks are very simple, this should not be a problem to implement the FDS in real world applications.
The financial support of the French Ministry of Foreign Affairs (under contract EGIDE ZENON 11999SD) and of the Cyprus Research Foundation (under contract KY-rA/0305/02) is greatly acknowledged.
[1] S. Kalogirou, G. Florides, S. Lalot, B. Desmet, Development of a Neural Network-Based Fault Diagnostic System, World Renewable Energy Congress IX and Exhibition, on CD-ROM, 19-25 August 2006, Florence, Italy
[2] S. Kalogirou, S. Lalot, G. Florides, B. Desmet, Development of a Neural Network-Based Fault Diagnostic System for Solar Thermal Applications, Solar Energy, Vol. 82, No. 2, pp. 164-172, 2007.
[3] S. Lalot, O. P. Palsson, G. R. Jonsson, B. Desmet, 2007, Comparison of neural networks and Kalman filters performances for fouling detection in a heat exchanger, International Journal of Heat Exchangers
[4] S. Lalot, 2006, On-line detection of fouling in a water circulating temperature controller (WCTC) used in injection moulding, Part 1: principles, Applied Thermal Engineering, Volume 26, Issues 11-12, August 2006, Pages 1087-1094
[5] S. Lalot, 2006, On-line detection of fouling in a water circulating temperature controller (WCTC) used in injection moulding, Part 2: application, Applied Thermal Engineering, Volume 26, Issues 11-12, August 2006, Pages 1095-1105
[6] M. M. Prieto, J. M. Vallina, I. Suarez, and I. Martin, 2000, Application of a design code for estimating fouling on-line in a power plant condenser refrigerated by seawater, Proceedings of the ASME-ZSITS International Thermal Science Seminar, Bled (Slovenia), on CD-ROM
[7] U. Jordan, K. Vajen, Realistic Domestic Hot-Water Profiles in Different Time Scales, 2001, available at http://sel. me. wisc. edu/trnsys/trnlib/iea-shc-task26/iea-shc-task26-load-profiles-description-jordan. pdf
[8] I. Knight, N. Kreutzer, M. Manning, M. Swinton, H. Ribberink, 2007, European and Canadian non - HVAC Electric and DHW load profiles for use in simulating the performance of residual cogeneration systems. A report of subtask A of FC+COGEN-SIM the simulation of building integrated fuel cell and other cogeneration systems. Annex 42 of the International Energy Agency energy conservation in buildings and community systems programme
[9] S. Lalot, 2000, identification of the time parameters of solar collectors using artificial neural networks, EuroSun (ISES Europe Solar Congress), Copenhagen (Denmark), June 19-22
[10] M. Bosanac, S. Jensen, In-situ solar air collector array test, Solar Energy laboratory, Danish Technological Institute, December 1997
