Learning ensembles of population dynamics models and their application to modelling aquatic ecosystems

• A novel approach for learning ensembles of process-based models.
• We identify the design choices for learning such ensembles.
• We apply the approach on modelling of population dynamics in aquatic ecosystems.
• For predictive modelling tasks, ensembles perform better than single models.

Ensemble methods are machine learning methods that construct a set of models and combine their outputs into a single prediction. The models within an ensemble can have different structure and parameters and make diverse predictions. Ensembles achieve high predictive performance, benefiting from the diversity of the individual models and outperforming them.In this paper, we develop a novel method for learning ensembles of process-based models. We build upon existing approaches to learning process-based models of dynamic systems from observational data, which integrates the theoretical and empirical paradigms for modelling dynamic systems. In addition to observed data, process-based modelling takes into account domain-specific modelling knowledge.We apply the newly developed method and evaluate its utility on a set of problems of modelling population dynamics in aquatic ecosystems. Data on three lake ecosystems are used, together with a library of process-based knowledge on modelling population dynamics. Based on the evaluation results, we identify the optimal settings of the method for learning ensembles of process-based models, i.e., the optimal number of ensemble constituents (25) as well as the optimal way to select (using a separate validation set) and combine them (using simple average). Furthermore, the evaluation results show that ensemble models have significantly better predictive performance than single models. Finally, the ensembles of process-based models accurately simulate the current and predict the future behaviour of the three aquatic ecosystems.