To produce predictions of lake water quality, it is important that the lake thermal structure is simulated by a lake physical model and that the output from this model is reliable and accurate. The accuracy of the model output is dependent in part on the quality of the meteorological input data. Forecasts of water quality would also use modelled weather data, so the model performance with modelled weather data would also be important to know. At the Lough Feeagh site in Mayo, Ireland, there are three different meteorological datasets available. There are two weather stations at Lough Feeagh; one on the lake shoreline, the Newport Automatic Weather Station (NAWS) and one on the Automatic Water Quality Monitoring Station (AWQMS) in the lake, both examples of data which could typically be used to calibrate a model. The European Centre for Medium-Range Weather Forecasts (ECMWF) produce global numerical weather forecasts for users worldwide. These data are available at a resolution of 0.125 x 0.125° (~256km2) at 3 hourly intervals for the Burrishoole catchment (Dee et al., 2011), and represent a typical forecast dataset for the grid square which includes Lough Feeagh
Daily values for the meteorological datasets from these three sources were collated for two years (2013-2014) and prepared as input files to provide forcing data for GOTM (the General Ocean Turbulence Model). Interpolated daily temperature profiles from 12 thermistors at depths of 2.5m, 5m, 8m, 11m, 14m, 16m, 18m, 20m, 22m, 27m, 32m and 42m from the AWQMS were used to validate the model. Figure 1 shows the observed temperature profile with the interpolated values. The model was run with each of the three meteorological datasets to simulate a daily temperature profile of the lake The models were then calibrated manually for each of the forcing datasets. Statistical analysis was then carried out to compare the differences between the model runs.
The main forcing data variables that affect model performance are air temperature, wind speed and mean sea level pressure. These were compared between the three datasets, using a correlation coefficient as a measure of association. The AWQMS and the NAWS were highly correlated across all three variables, with the lowest correlction value being for wind speed (Table 1). The correlations between the ECMWF data and the other two datasets had lower values with again that for wind speed being the lowest .
Table 1. Correlations between variables within the datasets.
|Air Temperature||Mean Sea Level Pressure||Wind Speed|
|ECMWF vs. NAWS||0.915||0.998||0.813|
|ECMWF vs. AWQMS||0.943||0.995||0.810|
|NAWS vs. AWQMS||0.979||0.998||0.890|
The calibration of the model with each dataset optimised the efficiency of the model. The factor that varies the most was the wind factor which was 0.834 using the ECMWF data and 1.22 for both the NAWS and AWQMS (Table 2).
The modelled output using the AWQMS forcing data showed deviation from the observed data with 20% of the simulations being within ± 0.1°C of the observed lake temperature values. The range of this deviation was -2.8°C to 3.4°C. 12% of the simulations using the ECMWF forcing data were within ± 0.1°C of the observed water temperature values, with a range of -2.8 to 2.6°C. The values for the NAWS forcing data were 20%, with a range of -2.2 to 2.8°C (Fig 2). While the ECMWF had the lowest deviation range, it consistently overestimated temperatures in winter and underestimated temperatures in the hypolimnion in summer. It was also notable that the simulations using the modelled ECMWF forcing data overestimated the temperatures in the surface waters (Fig. 2). All datasets underestimated the hypolimnetic temperatures in the summer but the AWQMS underestimated it the least (Fig 2). The ranges in the hypolimnetic temperatures were -2.77 to 0°C for the AWQMS, minus 2.79 to 0°C for the ECMWF and minus 2.25 to 0°C for the NAWS. The Nash-Sutcliffe efficiency of each model run at each depth throughout the water column is shown in Fig 3. All three simulations models had high values (>0.88) with the AWQMS having higher values in the upper depths of the lake (< 18m) while models have similar values in the lower depths of the lake (26-44m).
Table 2. Calibrated values of factors in the model.
|Swr_factor (Shortwave radiation)||Shf_factor (surface-heat flux)||Wind_factor||K_min (minimum turbulent kinetic energy)|
Table 3. Overall measures of model efficiency for each of the datasets – mean absolute error (MAE), root mean squared error (RMSE), Nash-Sutcliffe Efficiency (NSE), correlation (Cor), standard deviation of the model (s.d. mod) and standard deviation of the observed values (s.d. obs).
|MAE||RMSE||NSE||Cor||s.d. mod||s.d obs|
Fig 2. Temperature difference between the modelled water column temperature and the observed water column temperature with (a.) the ECMWF dataset, (b.) the NAWS and (c.) the AWQMS. Red represents areas where the model overestimates the temperature, blue represents areas where the model underestimates the temperature and white represents here the model is within ± 0.1°C of the observed temperature.
Fig 3. Nash-Sutcliffe Efficiency throughout the water column.
The meteorological dataset from the AWQMS produced the best model output based on the Nash-Sutcliffe values through the water column, and the overall model statistics. This is as might be expected and most likely related to the air temperatures being measured at the lake monitoring station being more representative of the air temperatures affecting the lake temperatures. Also the wind speed measured here would be accurately measure at the AWQMS than at the NAWS. This improved the accuracy of the model. However, the ECMWF forcing data, which was taken to represent a modelled weather forecast dataset, also performed well with values that were close to those of the data measured on the lake monitoring station.