A hybrid automaton is an appropriate mathematical formalism for modelling systems with mixed discrete and continuous dynamics. The hybrid automaton combines discretecontrol graphs with continuous dynamics defined by differential equations. In this paper, we introduce an offlinelearning algorithm to automatically synthesise a hybrid automaton from time series. The algorithm consists of severalprocedures, including segmentation of time series, structurelearning, and discovery of both dynamics and conditions.We present a novel method for determining the conditionsof a model learned from stationary time series, as well as anew time-based method for models learned from time seriesthat exhibit trends.The evaluation shows that the algorithm can learn simplemodels with an accurate graph structure and appropriatedynamics from stationary time series using variable conditions, as well as from non-stationary time series exhibitingtrends using the timed conditions. However, the results showpoor performance if the change point detection algorithmis unable to accurately segment the time series, or if thelocation dynamics are indistinguishable. Indistinguishabledynamics between locations may cause the change pointdetection to misidentify segments, and it may result in thestructure learning procedure being unable to identify thelocations as well as to associate the segments to locations. A hybrid automaton is an appropriate mathematical formalism for modelling systems with mixed discrete and continuous dynamics. The hybrid automaton combines discretecontrol graphs with continuous dynamics defined by differential equations. In this paper, we introduce an offlinelearning algorithm to automatically synthesise a hybrid automaton from time series. The algorithm consists of severalprocedures, including segmentation of time series, structurelearning, and discovery of both dynamics and conditions.We present a novel method for determining the conditionsof a model learned from stationary time series, as well as anew time-based method for models learned from time seriesthat exhibit trends.The evaluation shows that the algorithm can learn simplemodels with an accurate graph structure and appropriatedynamics from stationary time series using variable conditions, as well as from non-stationary time series exhibitingtrends using the timed conditions. However, the results showpoor performance if the change point detection algorithmis unable to accurately segment the time series, or if thelocation dynamics are indistinguishable. Indistinguishabledynamics between locations may cause the change pointdetection to misidentify segments, and it may result in thestructure learning procedure being unable to identify thelocations as well as to associate the segments to locations.