Exploring seasonal patterns using process modelling and evolutionary computation

Process models of chlorophyll-a concentration for freshwater systems, and in particular lake environments, have been developed over many years. Previous work has demonstrated that the optimisation of constants within these models has been able to significantly improve the quality of the resulting model on unseen data. This paper explores two properties of one particular process model: can the model predictions be improved by optimising the constants over different temporal scales; and can seasonal patterns be identified, based on monthly training scales, that allow further understanding of the response of the freshwater system to changing environmental conditions. The results show that there is some improvement on the prediction of unseen data when using constants of the process model optimised for individual months, versus constants trained over a yearly cycle. Additionally, by studying the patterns of the constants over various time scales some underlying seasonal patterns can be observed. These patterns can be further studied by exploring how the various elements of the process model vary with monthly versus yearly training constants. This work demonstrates some possible directions for understanding how the behaviour of freshwater systems at different time scales can be used to understand the properties of these complexes, non-linear systems. The results also suggest that local models of ecological time series data can be used to extract information that may not be obtainable from a single, global model.