Monitoring, fault detection and operation prediction of MSW incinerators using multivariate statistical methods Gilberto Tavares a , Zdena Zsigraiová b,1 , Viriato Semiao a,⇑ , Maria da Graca Carvalho a a Department of Mechanical Engineering, Instituto Superior Tecnico, Technical University of Lisbon, Av. Rovisco Pais, 1049-001 Lisbon, Portugal b Department of Furnaces and Thermal Technology, Technical University of Košice, Letná 9/A, 042 00 Košice, Slovakia article info Article history: Received 30 July 2010 Accepted 4 February 2011 Available online 4 March 2011 abstract This work proposes the application of two multivariate statistical methods, principal component analysis (PCA) and partial least square (PLS), to a continuous process of a municipal solid waste (MSW) moving grate-type incinerator for process control – monitoring, fault detection and diagnosis – through the extraction of information from historical data. PCA model is built for process monitoring capable of detecting abnormal situations and the original 16-variable process dimension is reduced to eight, the first 4 being able to capture together 86% of the total process variation. PLS model is constructed to predict the generated superheated steam flow rate allowing for control of its set points. The model retained six of the original 13 variables, explaining together 90% of the input variation and almost 98% of the output varia- tion. The proposed methodology is demonstrated by applying those multivariate statistical methods to process data continuously measured in an actual incinerator. Both models exhibited very good perfor- mance in fault detection and isolation. In predicting the generated superheated steam flow rate for its set point control the PLS model performed very well with low prediction errors (RMSE of 3.1 and 4.1). Ó 2011 Elsevier Ltd. All rights reserved. 1. Introduction After the entry into force of the European Directive on waste landfilling (EC, 1999) incineration with energy recovery became common practice in some European countries (ISWA, 2002) as the final disposal for the huge amounts of unsorted domestic waste daily generated. Waste-to-energy (WTE) conversion has also been the unsorted MSW final disposal in countries with severe availabil- ity constraints on land resources (Mastro and Mistretta, 2004; Cap- uto et al., 2004; Luoranen and Horttanainen, 2008). Moreover, the increasingly competitive global market and the compliance with the very strict emissions limits imposed by legis- lation require higher energy efficiencies and lower emissions in the environment for the WTE incinerators operation. Such target can be successfully achieved through the use of on- line process fault detection and diagnosis tools (Leskens et al., 2005) that can avoid abnormal events progression and to maintain equipments operation close to their design conditions. This is par- ticularly relevant for mass-burning incinerators due to the fre- quent and unpredictable change in waste composition. Fault diagnosis requires a priori knowledge of relationships be- tween symptoms and failures (Venkatasubramanian et al., 2003a,b). Such knowledge can be extracted from past experience, i.e. historical operation data of the incineration process through the use of multivariate statistical methods (Geladi and Kowalski, 1986; Venkatasubramanian et al., 2003c). Projection methods like PCA (principal component analysis) and PLS (partial least square) have increasingly gained importance in extracting essential infor- mation from huge amounts of original data. Initially proposed by Pearson (1901) and later developed by Hotteling (1947), PCA method is a standard multivariate projection technique applied to a matrix of process variables aiming at reduc- ing the problem dimension and extracting information. Conceptually similar, PLS is useful in reducing simultaneously the dimensions of the process (independent) variables and the cor- responding response (dependent) variables (Jackson, 1991; Yoon and MacGregor, 2001). Both PCA and PLS models are built from data collected during normal operation. Detection of any abnormal behaviour is per- formed through deviations between the model predictions and the projected measurements. The use of PCA and PLS methods to analyse and monitor processes is well established (Geladi and Kowalski, 1986; Jackson, 1991; Kourti and MacGregor, 1995; Rus- sell et al., 2000), particularly in the chemical industry (Kourti et al., 1996; Wise and Gallagher, 1996; Lee et al., 2004). Applications for monitoring and detecting abnormal situations over time range from batch and continuous processes (Chiang et al., 2000; Lee 0956-053X/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.wasman.2011.02.005 ⇑ Corresponding author. Tel.: +351 841 7726. E-mail addresses: gtavares@ist.utl.pt (G. Tavares), zdena@ist.utl.pt (Z. Zsi- graiová), ViriatoSemiao@ist.utl.pt (V. Semiao), Maria.carvalho@ist.utl.pt (Maria da Graca Carvalho). 1 Presently at IDMEC/IST, Lisbon, Portugal. Waste Management 31 (2011) 1635–1644 Contents lists available at ScienceDirect Waste Management journal homepage: www.elsevier.com/locate/wasman