Abstract
Industrial activities fall under the scope of the Seveso Directive as a function of the amount of dangerous substances present in the site. In this sense, as defined in point 12 of Art. 3 of the New Seveso III Directive, “the ‘presence of dangerous substances’ means the actual or anticipated presence of dangerous substances in the establishment, or of dangerous substances which it is reasonable to foresee may be generated during loss of control of the processes, including storage activities, in any installation within the establishment”.
Experience shows that the requirement of declaring the presence of dangerous substances due to a loss of control of an industrial chemical process is hard to fulfil; in first place due to the complexity of defining a loss of control scenario and secondly because of the lack of information on the substances that may be generated when processes do not follow the desired reactive behaviour.
Since the behaviour of chemical substances on a chemical reaction does not follow a random path, this is an interesting and suitable field where to apply data mining techniques. These techniques intend to identify empirical regularities observed over a large data set, which are believed to be useful with substance prediction purposes.
This paper explores different possible techniques that can be applied with the purpose of predicting the substances that may be generated under loss of control conditions. Suitability of different techniques is evaluated by measuring its accuracy on the prediction of known scenarios.
