Using physically estimated temperatures as inputs to enhance the performance of an LSTM model for predicting water temperatures in fishponds

• The physics-augmented LSTM model greatly improves short-term pond temperature prediction in a fishpond. • The simulated water temperature input reduces errors and lowers the data dependency in the LSTM model. • Accurate forecasts support proactive aquaculture management and reduce weather-related losses. Accurate forecasting of water temperature is critical for sustainable aquaculture management, especially given the growing challenges posed by climate variability. In this study, a novel hybrid modeling approach that integrates a one-dimensional heat energy physical model with a long short-term memory (LSTM) neural network to forecast fishpond water temperature is introduced. The hybrid LSTM–physical model combines the strengths of physically based simulations and data-driven learning, which enables it to capture both fundamental thermal processes and complex temporal patterns. A key innovation of this approach is to evaluate whether incorporating physically simulated water temperature—estimated using the physical model driven by meteorological inputs such as solar radiation, air temperature, wind speed, and relative humidity—can enhance the generalization and predictive accuracy of an LSTM model. Model performance was validated using observational data from a fishpond in Hsinchu, Taiwan, collected between November 10, 2023, and April 18, 2024. The results indicate that the hybrid LSTM–physical model consistently outperforms the standard LSTM model, which relies solely on meteorological inputs. On average, it achieves average improvements of 48.65% in root mean square error (RMSE), 27.12% in the coefficient of determination (R²), and 63.07% in mean bias error (MBE) across all forecast horizons (1, 2, and 3 hours ahead). These findings suggest that incorporating physically simulated temperature enables the model to capture underlying system dynamics that purely data-driven approaches may overlook or require substantially more data to learn effectively.