Abstract: Ecological forecasting is emerging as a valuable tool for generating predictions, with quantified uncertainty, to aid management in the face of increased ecosystem variability due to land use and climate change. On a near-term scale (days to months ahead), ecological forecasts provide valuable information that can inform preemptive decision-making to improve resource management and limit or prevent ecosystem degradation. In freshwater ecosystems, water temperature forecasts are especially needed to guide water quality management, drinking water extraction, and dam releases for optimising downstream temperatures, but the optimal modelling framework for producing forecasts remains unknown. Specifically, there is still debate as to how to best use imperfect models that are suboptimal at various spatial and temporal scales to inform decision making. One possible approach for improving forecast performance is through multi-model ensemble (MME) forecasts, which integrate information and uncertainty from multiple models, potentially improving forecast skill over more scales. This approach has proved to be successful in other fields, but it has yet to be implemented for lake water temperature forecasting.
Here, we integrated three process-based models and two empirical models to generate ensemble forecasts of water temperature in a drinking water supply reservoir. We demonstrate the promise of MMEs in improving water temperature forecast skill over near-term horizons. Across all forecast horizons between 1 and 14 days ahead, the integrated MME exhibited 10-60% improvement over the skill of any one individual model, assessed using a probabilistic scoring metric. The model generating the best forecast at 1 day ahead was different from the best forecast at 14 days ahead, highlighting the need for different models at different horizons. Our results show that no individual model performed best over all time scales or depths, demonstrating that the MME can provide a robust forecast that is not possible if only one model is used. Moreover, we also show that baseline empirical models (specifically, persistence and day of year mean models) provide additional useful forecasting information that process models may lack. By integrating models from two contrasting model types (process-based and empirical), we found that the water temperature forecast skill improved by 10% compared to an ensemble composed of just process models. Altogether, our work informs water temperature forecasts for lake management as well as the future development of ecological forecasting methods across multiple ecosystems and ecological variables.
