LAUSR

Forecasting water quality in inland waters can improve management practices to protect water resources. This study proposes a novel data‐driven framework to forecast water quality profiles over long time periods in Boulder Basin of Lake Mead, a deep monomictic subtropical lake. Hourly meteorological data were used to estimate lake–atmosphere heat exchange. Heat fluxes combined with 6‐hourly measured water quality profiles up to 106 m depth were used to train six different artificial neural networks to forecast water temperature, dissolved oxygen, and conductivity profiles up to 240 d ahead. A model incorporating heat fluxes, winds, and stationary wavelet decomposition generated correlation coefficients > 0.88 and relative errors < 4% throughout the water column for up to 240‐d ahead forecasts. Internal wave motions at the thermocline resulted in larger relative errors of forecasts in the metalimnion compared to other depths. Greater atmospheric influences on water temperature and dissolved oxygen resulted in larger forecast errors compared to conductivity. An autocovariance method successfully determined appropriate forecasting lead times at different depths, improving forecast accuracies.