Mon, Aug 15, 2022: 2:00 PM-2:15 PM
513A
Background/Question/MethodsThere is a growing interest in developing near-term forecasts of the carbon and water cycles. Most such forecasts use process-based models to predict pools and fluxes, but because all models are simplifications of reality there will often be systematic high-dimensional patterns to model errors. Here we investigate whether near-term probabilistic forecasts of carbon and water pools and fluxes generated using the Simplified Photosynthesis and Evapotranspiration Model (SIPNET) can be improved by using flexible statistical modeling approaches to identify and correct for errors in model outputs. We developed a suite of Bayesian models to predict errors in both forecast accuracy and precision for net ecosystem exchange (NEE) and soil moisture across ten eddy covariance sites using time-of-day, day-of-year, SIPNET model output and historical means of soil moisture and NEE as inputs. These models were calibrated against forecasts from 2019-2021 and validated using forecasts from 2022.
Results/ConclusionsWe hypothesized that as we added more covariates to the statistical model the forecast RMSE and CRPS scores would decrease and that the addition of the historical means would have the largest bias correcting effect. We found that the bias in predictions of NEE and soil moisture was most improved through addition of historical means. We also found that forecast errors were autocorrelated over short timescales and that the forecast variance for NEE was highly heteroskedastic, requiring corrections to the predictive spread across both the diurnal and seasonal cycles to achieve proper probabilistic coverage. Deciduous forests had a larger decrease in model bias compared to evergreen forests and non-woody ecosystems. Building upon this initial analysis, we will further expand our statistical model to include meteorological covariates, such as temperature, precipitation, and solar radiation, and focus on identifying model structural errors.
Results/ConclusionsWe hypothesized that as we added more covariates to the statistical model the forecast RMSE and CRPS scores would decrease and that the addition of the historical means would have the largest bias correcting effect. We found that the bias in predictions of NEE and soil moisture was most improved through addition of historical means. We also found that forecast errors were autocorrelated over short timescales and that the forecast variance for NEE was highly heteroskedastic, requiring corrections to the predictive spread across both the diurnal and seasonal cycles to achieve proper probabilistic coverage. Deciduous forests had a larger decrease in model bias compared to evergreen forests and non-woody ecosystems. Building upon this initial analysis, we will further expand our statistical model to include meteorological covariates, such as temperature, precipitation, and solar radiation, and focus on identifying model structural errors.