Modulations, the moments where key change, are structurally important in tonal music. Analyzing music, especially studying large-scale music structure of a piece, often implies to look for modulations. State-of-the-art keyfinding algorithms generally aim at identifying keys rather than studying the way they change. Here, we introduce new ways to model modulations with the help of features based on musicological knowledge, as well as an algorithm estimating the tonal plan of a piece. We study the concept of current diatonic pitch set and introduce a heuristic to detect dominant-to-tonic progressions. We design three proximity measures to assess how close the music is from each key. These measures are then combined by an algorithm that identifies an optimal tonal plan. We report results on a corpus including 38 movements from Mozart’s string quartets, obtaining a 84.8% prediction of correct keys with insight on where the modulations occur.