Forecasting surface water temperature in lakes: A comparison of approaches

Accurate water temperature forecasting in lake systems is important for environmental impact assessment and fisheries management, among others. In this study, two models are developed and applied for water temperature forecasting in lake systems: (1) the multi-layer perceptron neural network (MLPNN) model; and (2) the wavelet transform and MLPNN integrated model (WT_MLPNN). The models are applied to forecast daily lake surface water temperature (LSWT) of eight lowland Polish lakes. Long-term daily LSWT from eight lakes and daily air temperatures from seven meteorological stations are used for daily LSWT forecasting. The results from the two models are compared with those obtained from two other widely used models: the physically-statistically based hybrid air2water model and a non-linear regression model (S-curve). The modelling results show that the air2water model performs the best, followed by the WT_MLPNN, MLPNN, and the non-linear regression model. Overall, the air2water, WT_MLPNN, and MLPNN models reproduce well the seasonal and inter-annual variations of the LSWT dynamics in the eight lakes. The non-linear regression model, although providing the lowest accuracy, can still provide good preliminary estimates of the LSWT for the eight lakes. The outcomes of the present research can provide references for forecasting lake surface water temperature and sustainable management of lake ecosystems.