The high proportion of new energy integration into the basin’s cascaded hydropower system significantly increases the uncertainty of grid scheduling. To address the highly coupled, nonlinear, multi-constraint, and multi-objective optimization challenges faced by the cascaded hydro-wind-photovoltaic complementary system, an optimal day-ahead scheduling method based on a physically-embedded neural network was proposed. In the modeling phase, core physical mechanisms, such as the turbine efficiency curve, reservoir water balance equation, and wind power probability distribution, were embedded as physical constraint loss terms and integrated into a fully connected neural network architecture. This integration was aimed at enhancing the interpretability of scheduling outcomes and overcoming the limitations of traditional purely data-driven methods. During the training phase, by leveraging the DeepXDE framework, the physical constraint loss term and the data-driven loss term were jointly optimized. An adaptive weight adjustment mechanism was introduced to coordinate multi-source information between historical operational data and physical constraint equations, thereby improving the physical plausibility of the scheduling results and their consistency with actual operations. Furthermore, an adaptive weighting optimization strategy for multi-source losses was adopted to mitigate numerical instability and convergence difficulties caused by gradient magnitude discrepancies during training, thereby facilitating the generation of optimal scheduling plans. Comparative validation using an actual basin dataset demonstrated the model’s validity and generalization capability. It is shown to produce generation scheduling strategies that balance system reliability and economy, significantly enhancing the system’s robustness in coping with fluctuations in wind and solar power, as well as the dynamic synergistic optimization performance of hydro-wind-solar multi-energy gradient utilization.