No abstract

No abstract

Conventional Life Cycle Assessment (LCA) that relies on static coefficients is usually based on yearly averages. However, the impacts of electricity supply vary remarkably on an hourly basis. Thus, a company production plan is reassessed to reduce selected LCA impacts due to electricity consumption. To achieve this, the company will need a forecast of hourly LCA impacts due to electricity consumption, which can be directly forecast with the Direct Forecasting (DF) approach. Alternatively, the Electricity Technological Mix Forecasting (ETMF) forecasts the electricity production of the technologies in the mix and subsequently linearly combines it with unitary LCA impact indicators. Here, we assessed different machine learning models to forecast two LCA impact indicators for the consumption of electricity in the Italy-North control zone. The feed-forward neural network (NN) with the ETMF approach was the best perfomer among the assessed forecasting models. In our dataset, recurrent neural networks (RNNs) performed worse than feed-forward neural networks. Due to its better forecasting performance, the ETMF approach was preferred over the DF approach. This was due to its flexibility and scalability with easy updates or expansion of the selected forecast indicators, and due to its ability to assess technology-specific errors in the forecasting. Finally, we propose to adopt the correlation of LCA impact indicators within the dataset to select indicators while avoiding unconscious burden-shifting.

No abstract

The automatic extraction of a process model from data is one of the main focuses of a Process Mining pipeline. Decision Mining aims at discovering conditions influencing the execution of a given process instance to enhance the original extracted model. In particular, a Petri Net with data is a Petri Net enhanced with guards controlling the transitions firing in correspondence of places with two or more output arcs, called decision points. To automatically extract guards, Decision Mining algorithms fit a classifier for each decision point, indicating what path the case will follow based on event attributes. Retrieving the path followed by the case inside the model is crucial to create each decision point's training dataset. Indeed, due to the presence of invisible activities, having multiple paths coherent with the same trace in the event log is possible. State-of-the-art method consider only the optimal path discarding the other possible ones. Consequently, training sets of related decision points will not contain information on the considered case. This work proposes a depth-first-based method that considers multiple paths possibly followed by a case inside the Petri Net to avoid information loss. We applied the proposed method to a real-life dataset showing its effectiveness and comparing it to the current state of the art.