<p>Traditional pond aquaculture water quality prediction models have limitations when processing cross-source heterogeneous data, particularly due to their failure to fully account for the impact of future meteorological data on water quality changes. Meteorological factors like temperature, air pressure, and rainfall typically cause significant lag effects on water quality changes. Existing models often rely solely on historical water quality data for predictions, overlooking the influence of meteorological factors. This paper proposes an enhanced deep learning model, the dual encoder cross-source feedback network (DECSF-Net), incorporating a modified dual encoder structure to encode water quality and meteorological data separately. This design accurately captures the complex impact of future meteorological data on water quality time series. The cross-source feedback fusion (CSFF) module enhances mutual attention between water quality and meteorological data through a bidirectional feedback mechanism, improving the model’s ability to jointly represent cross-source data. Experimental results demonstrate that DECSF-Net outperforms existing mainstream methods in predicting water quality for the next 8&#xa0;h, with a mean squared error (MSE) of 0.0959, mean absolute error (MAE) of 0.2037, root mean squared error (RMSE) of 0.3084, and mean absolute percentage error (MAPE) of 1.5159, showcasing its superior prediction accuracy. This model effectively addresses the water quality prediction challenges in complex ecological environments. The paper shows that integrating future meteorological data into water quality prediction methods significantly improves accuracy, offering substantial practical value.</p>